A script for making patches

I have a standard format for patchnames: 1234-99.project.brief-description.patch, where 1234 is the issue number and 99 is the (expected) comment number. However, it involves two copy-pastes: one for the issue number, taken from my browser, and one for the project name, taken from my command line prompt.

Some automation of this is clearly possible, especially as I usually name my git branches 1234-brief-description. More automation is less typing, and so in true XKCD condiment-passing style, I've now written that script, which you can find on github as dorgpatch. (The hardest part was thinking of a good name, and as you can see, in the end I gave up.)

Out of the components of the patch name, the issue number and description can be deduced from the current git branch, and the project from the current folder. For the comment number, a bit more work is needed: but now has a public API, so a simple REST request to that gives us data about the issue node including the comment count.

So far, so good: we can generate the filename for a new patch. But really, the script should take care of doing the diff too. That's actually the trickiest part: figuring out which branch to diff against. It requires a bit of git branch wizardry to look at the branches that the current branch forks off from, and some regular expression matching to find one that looks like a Drupal development branch (i.e., 8.x-4.x, or 8.0.x). It's probably not perfect; I don't know if I accounted for a possibility such as 8.x-4.x branching off a 7.x-3.x which then has no further commits and so is also reachable from the feature branch.

The other thing this script can do is create a tests-only patch. These are useful, and generally advisable on issues, to demonstrate that the test not only checks for the correct behaviour, but also fails for the problem that's being fixed. The script assumes that you have two branches: the one you're on, 1234-brief-description, and also one called 1234-tests, which contains only commits that change tests.

The git workflow to get to that point would be:

  1. Create the branch 1234-brief-description
  2. Make commits to fix the bug
  3. Create a branch 1234-tests
  4. Make commits to tests (I assume most people are like me, and write the tests after the fix)
  5. Move the string of commits that are only tests so they fork off at the same point as the feature branch: git rebase --onto 8.x-4.x 1234-brief-description 1234-tests
  6. Go back to 1234-brief-description and do: git merge 1234-tests, so the feature branch includes the tests.
  7. If you need to do further work on the tests, you can repeat with a temporary branch that you rebase onto the tip of 1234-tests. (Or you can cherry-pick the commits. Or do cherry-pick with git rev-list, which is a trick I discovered today.)

Next step will be having the script make an interdiff file, which is a task I find particularly fiddly.

A git-based patch workflow for (with interdiffs for free!)

There's been a lot of discussion about how we need github-like features on Will we get them? There's definitely many improvements in the pipeline to the way our issue queues work. Whether we actually need to replicate github is another debate (and my take on it is that I don't think we do).

In the meantime, I think that it's possible to have a good collaborative workflow with what we have right now on, with just the issue queue and patches, and git local branches. Here's what I've gradually refined over the years. It's fast, it helps you keep track of things, and it makes the most of git's strengths.

A word on local branches

Git's killer feature, in my opinion, is local branches. Local branches allow you to keep work on different issues separate, and they allow you to experiment and backtrack. To get the most out of git, you should be making small, frequent commits.

Whenever I do a presentation on git, I ask for a show of hands of who's ever had to bounce on CMD-Z in their text editor because they broke something that was working five minutes ago. Commit often, and never have that problem again: my rule of thumb is to commit any time that your work has reached a state where if subsequent changes broke it, you'd be dismayed to lose it.

Starting work on an issue

My first step when I'm working on an issue is obviously:

  git pull

This gets the current branch (e.g. 7.x, 7.x-2.x) up to date. Then it's a good idea to reload your site and check it's all okay. If you've not worked on core or the contrib project in question in a while, then you might need to run update.php, in case new commits have added updates.

Now start a new local branch for the issue:

  git checkout -b 123456-foobar-is-broken

I like to prefix my branch name with the issue number, so I can always find the issue for a branch, and find my work in progress for an issue. A description after that is nice, and as git has bash autocompletion for branch names, it doesn't get in the way. Using the issue number also means that it's easy to see later on which branches I can delete to unclutter my local git checkout: if the issue has been fixed, the branch can be deleted!

So now I can go ahead and start making commits. Because a local branch is private to me, I can feel free to commit code that's a total mess. So something like:

  // Commented-out earlier approach that didn't quite work right.
  $foo += $bar;
  // Badly-formatted code that will need to be cleaned up.
  if($badly-formatted_code) { $arg++; }

That last bit illustrates an important point: commit code before cleaning up. I've lost count of the number of times that I've got it working, and cleaned up, and then broken it because I've accidentally removed an important line that was lost among the cruft. So as soon as code is working, I make a commit, usually whose message is something like 'TOUCH NOTHING IT WORKS!'. Then, start cleaning up: remove the commented-out bits, the false starts, the stray code that doesn't do anything, in small commits of course. (This is where you find it actually does, and breaks everything: but that doesn't matter, because you can just revert to a previous commit, or even use git bisect.)

Keeping up to date

Core (or the module you're working on) doesn't stay still. By the time you're ready to make a patch, it's likely that there'll be new commits on the main development branch (with core it's almost certain). And before you're ready, there may be commits that affect your ongoing work in some way: API changes, bug fixes that you no longer need to work around, and so on.

Once you've made sure there's no work currently uncommitted (either use git stash, or just commit it!), do:

git fetch
git rebase BRANCH

where BRANCH is the main development branch that is being committed to on, such as 8.0.x, 7.x-2.x-dev, and so on.

(This is arguably one case where a local branch is easier to work with than a github-style forked repository.)

There's lots to read about rebasing elsewhere on the web, and some will say that rebasing is a terrible thing. It's not, when used correctly. It can cause merge conflicts, it's true. But here's another place where small, regular commits help you: small commits mean small conflicts, that shouldn't be too hard to resolve.

Making a patch

At some point, I'll have code I'm happy with (and I'll have made a bunch of commits whose log messages are 'clean-up' and 'formatting'), and I want to make a patch to post to the issue:

  git diff 7.x-1.x > 123456.PROJECT.foobar-is-broken.patch

Again, I use the issue number in the name of the patch. Tastes differ on this. I like the issue number to come first. This means it's easy to use autocomplete, and all patches are grouped together in my file manager and the sidebar of my text editor.

Reviewing and improving on a patch

Now suppose Alice comes along, reviews my patch, and wants to improve it. She should make her own local branch:

  git checkout -b 123456-foobar-is-broken

and download and apply my patch:

  patch -p1 < 123456.PROJECT.foobar-is-broken.patch

(Though I would hope she has a bash alias for 'patch -p1' like I do. The other thing to say about the above is that while wget is working at downloading the patch, there's usually enough time to double-click the name of the patch in its progress output and copy it to the clipboard so you don't have to type it at all.)

And finally commit it to her branch. I would suggest she uses a commit message that describes it thus:

  git commit -m "joachim's patch at comment #1"

(Though again, I would hope she uses a GUI for git, as it makes this sort of thing much easier.)

Alice can now make further commits in her local branch, and when she's happy with her work, make a patch the same way I did. She can also make an interdiff very easily, by doing a git diff against the commit that represents my patch.

Incorporating other people's changes to ongoing work

All simple so far. But now suppose I want to fix something else (patches can often bounce around like this, as it's great to have someone else to spot your mistakes and to take turns with). My branch looks like it did at my patch. Alice's patch is against the main branch (for the purposes of this example, 7.x-1.x).

What I want is a new commit on the tip of my local branch that says 'Alice's changes from comment #2'. What I need is for git to believe it's on my local branch, but for the project files to look like the 7.x-1.x branch. With git, there's nearly always a way:

  git checkout 7.x-1.x .

Note the dot at the end. This is the filename parameter to the checkout command, which tells git that rather than switch branches, you want to checkout just the given file(s) while staying on your current branch. And that the filename is a dot means we're doing that for the entire project. The branch remains unchanged, but all the files from 7.x-1.x are checked out.

I can now apply Alice's patch:

  patch -p1 < 123456.2.PROJECT.foobar-is-broken.patch

(Alice has put the comment ID after the issue ID in the patch filename.)

When I make a commit, the new commit goes on the tip of my local branch. The commit diff won't look like Alice's patch: it'll look like the difference between my patch and Alice's patch: effectively, an interdiff. I now make a commit for Alice's patch:

  git commit -m "Alice's patch at comment #2"

I can make more changes, then do a diff as before, post a patch, and work on the issue advances to another iteration.

Here's an example of my local branch for an issue on Migrate I've been working on recently. You can see where I made a bunch of commits to clean up the documentation to get ready to make a patch. Following that is a commit for the patch the module maintainer posted in response to mine. And following that are a few further tweaks that I made on top of the maintainer's patch, which I then of course posted as another patch.

A screenshot of a git GUI showing the tip of a local branch, with a commit for a patch from another user.

(Notice how in a local branch, I don't feel the need to type terribly accurately for my commit messages, or indeed be all that clear.)

Improving on our tools

Where next? I'm pretty happy with this workflow as it stands, though I think there's plenty of scope for making it easier with some git or bash aliases. In particular, applying Alice's patch is a little tricky. (Though the stumbling block there is that you need to know the name of the main development branch. Maybe pass the script the comment URL, and let it ask what the branch of that issue is?)

Beyond that, I wonder if any changes can be made to the way git works on A sandbox per issue would replace the passing around of patch files: you'd still have your local branch, and merge in and push instead of posting a patch. But would we have one single branch for the issue's development, which then runs the risk of commit clashes, or start a new branch each time someone wants to share something, which adds complexity to merging? And finally, sandboxes with public branches mean that rebasing against the main project's development can't be done (or at least, not without everyone know how to handle the consequences). The alternative would be merging in, which isn't perfect either.

The key thing, for me, is to preserve (and improve) the way that so often on, issues are not worked on by just one person. They're a ball that we take turns pushing forward (snowball, Sisyphean rock, take your pick depending on the issue!). That's our real strength as a community, and whatever changes we make to our toolset have to be made with the goal of supporting that.

Git tricks: repatching for an issue branch

My workflow for making patches is to use a feature branch for a single issue. Whether you're a contributor or a maintainer it lets you advance the fixing of the problem in small increments, and safely experiment knowing you can roll back.

But where it goes wrong is when your patch is superseded by a newer one in the issue queue, and you want to work on it some more. How do you update your branch for the ongoing work? As ever, with git there's a way.

Let's start with the basics first: you're making a feature branch to work on an issue. I tend to follow the naming pattern '123456-fix-all-the-bugs', but for this example I'll call it 'issue'.

// Make a new branch and switch to it.
$ git co -b issue
// Make lots of commits.
// Ready to make a patch:
$ git diff > 123456.project.issue.patch

(Note that if you can make your patch to show all your commits one by one, which can sometimes aid in making it clear what you're changing, but that's for another day.)

You've now got a patch which you're uploading to the issue queue, and your tree looks something like this:

* [issue] Last commit, ready to roll a patch!
* Fixed the foobar.
* Added a bizbax.
* [master]

Now someone else comes along to the issue queue, reviews your patch, and posts a new patch of their own. You in turn look at patch 2, and while it's an improvement, you think it needs still more work.

The problem is how to apply the patch to your repository. It won't apply to the tip of the issue branch, and if you checkout master, you can't get back to your issue branch. You can of course just discard your original issue branch, and create a branch issue2 for patch 2.

Or you can do this:

// Start on the issue branch.
// Stash any work in progress!
$ git stash
// Checkout just the *files* of master, while keeping the HEAD pointer on the
// issue branch.
$ git checkout master -- .
// This puts the files from master into the working tree, but keeps the index
// on the issue branch. In simpler terms, the reverse of patch 1 will appear
// staged (as git believes that your files *ought* to look like patch 1, but
// actually look like master).
// We want the index clean, so unstage everything:
$ git reset HEAD .
// Now apply the new patch.
$ patch -p1 < patch-2.patch
// Now commit this as patch 2.
// Remember to stash pop when you're done!

Because the working tree files (that is, the actual files on your system) look like the master branch, the patch applies cleanly. But because git still believes its on the tip of the issue branch, the commit you make goes on the tip of that branch, and the diff it records is effectively the interdiff between your patch-1 and the other contributor's patch-2. Your tree looks like this:

* [issue] Applied patch 2 from Ada Lovelace.
* Last commit, ready to roll a patch!
* Fixed the foobar.
* Added a bizbax.
* [master]

Result: you can now do more work on this branch, and make more commits, and when you're ready, diff against master to make patch-3, ready to upload to the issue queue.

Git tricks: being on the wrong branch

I often find that I'm in the middle of one thing when I have to do another. Whether it's hotfixes for a client, or just finding a minor bug that blocks my current work, or needing to add components to a feature before I can add custom functionality.

The best way is to stash your current work, checkout the master branch, commit, then go back. If you're working on a feature branch (and you should be), then rebase that afterwards so you have access to the new work there. So that's:

$ git stash
$ git checkout master
// do commits
$ git checkout feature
$ git rebase master
$ git stash pop

But that's not always feasible. Sometimes I'm sloppy, and I've already made code changes before stashing. And lately, I've got one instance of Party module that's got a feature branch that's made database changes, but I don't want to hold that up ongoing commits (and I'm too lazy to set up a new local site!).

If your fix is just one commit, you can make it on the feature branch, then cherrypick it to the master branch like this:

// make your commit and note its SHA
$ git stash
$ git checkout master
$ git cherry-pick COMMIT
$ git checkout feature
$ git rebase master

The rebase should be smart enough to figure out the same commit exists on both branches, and will silently drop it from the feature branch.

Alternatively, if you want to do a chunk of main branch work, make a temporary branch on the tip of feature, which you can then move to the master branch when you're done:

$ git checkout -b moveme
// make as many commits as you like
// Now we take everything that's between the tips of feature and moveme, and move it to the tip of master
$ git rebase --onto master feature moveme
// Now merge moveme into master: this'll just fast-foward master.
$ git checkout master
$ git merge moveme
// moveme can be deleted now
$ git branch -d moveme
// Now rebase the feature branch
$ git checkout feature
$ git rebase master

As I've become more familiar with git, I've found that temporary, throw-away branches can be useful in a variety of situations. Another one is making a backup branch prior to potentially messy rebases: just create a branch 'backup' where you are to be sure that no matter what happens with the rebase, your current chain of commits will be preserved. If there are conflicts, you can diff against it to check no code was lost. And when you're happy with the result of the rebase, just delete it. Branches being cheap, and local, opens up a whole new set of uses for them.

Get out, git!

There are lots of good reasons to have your server's codebase be an actual git checkout. But there's one potential flaw: your entire repository's history ends up in your webroot inside a .git folder.

You can block access to it in your .htaccess, but that's hacking core (until this patch lands at least).

There is however an alternative method that lets you keep the entirety of git's working folder outside the webroot completely.

Here's how to convert an existing repository to this format:

  1. Move the .git folder to another location, renaming it in the process so it's no longer hidden. The convention is to leave it with a .git ending though, so for example, 'mysite.git'. I put these inside a folder called 'git' in the user's home folder, for instance.

    $ mv .git ~/git/mysite.git

  2. In its original place in your webroot, create a new file called '.git'. Into this file place a single line thus:

    gitdir: /absolute/path/to/your/mysite.git

    This needs to be an absolute path; relative ones confuse git when you go into subfolders. Using '~' to start at the user's home folder doesn't seem to work either.

  3. Finally, we need to tell the config file where the work folder is. This step isn't completely necessary, but it allows you to invoke the git command while standing in subfolders of your webroot, which is too handy a thing to lose.

    Standing either in the webroot or in the git folder, do:

    $ git config core.worktree "/absolute/path/to/your/webroot"

    You can also edit the git config file by hand to set this, which allows you to also add a comment explaining the manoeuvre for future reference.

That's all there is to it. You now have a working git repository whose working folder is completely inaccessible from the outside world.

For creating a new repo, you can use the following finger-twister:

$ git --git-dir=/path/to/repo.git --work-tree=. init && echo "gitdir: /path/to/repo.git" > .git

There are more tips in this question on StackOverflow. And for a hands-on tutorial, come to my session on git at DrupalCamp Scotland, taking place later this month in Edinburgh.

Moving a git local branch from one local to another

You're maybe at one of the many Drupal Co-worker Friday events that are taking place around the world today. You've packed up your laptop and your lunchbox, and you're looking forward to a day out of the house with some human contact.

But yesterday you were halfway through a big piece of work on your project. And you were using a git local branch, of course. Why? Because it keeps your work isolated off the main development branch, allowing other work to continue independently. And because while your commits are only local you're free to reorder them, edit the log messages, fixup mistakes as if they never happened, and so on. In fact, if you so choose, merging your local branch in with the --squash option makes it look like you made all of the work in a single, perfect commit. Wow!

But that branch is on your desktop machine, and today you're on the laptop. How to get it over from one to the other? You don't want to push it to the remote, because then that means you can't rework it, and it doesn't really belong there anyway. You could make one big patch of your branch against the development branch, but then on your laptop you won't have the dozen or so commits you made yesterday (and you work with small, simple commits, because it makes it easier to roll back should you need to, and to see what's been changed). Furthermore, you've a bunch of changes that aren't yet committed because they're work in progress. You need those too, but not mixed in with what's already committed and reasonably stable.

The solution is the git format-patch command. At first try this is a rather weird one, which fills your folder up with cryptic mailbox files (which as far as I can tell are nigh on useless in the modern world of webmail). But it has an option which turns it into something very useful indeed: --stout. So much so that I've added it to my git global config thus:

  fp  = format-patch --stdout

So standing in your repository and doing

  git fp the-dev-branch > my-local-branch.patch

will give you one single file that comprises multiple patches, one for each commit. Copy that to your laptop by your favourite means, and over there do:

  git co -b my-local-branch
  git am  my-local-branch.patch

And all your commits from your desktop machine are reproduced on your laptop.

But what about your work in progress? I was coming to that. Before you begin, make one commit of all of them, perhaps with a log message to remind you it's the work in progress.

Then after you've done 'git am', all we have to do is kill off that final tip commit, while keeping the changes it contains in the filesystem. The command for that is git reset with the --mixed option:

  git reset --mixed HEAD^

All your work in progress changes are now 'unstaged changes', and your laptop's copy of the repository is in exactly the same stage as your desktop machine.

What about going back the other way? Well I'll figure that one out tonight, I expect ;)

PS. The return trip is easily done thus:

When you create the branch on the new machine, but before you apply the patch, add a tag: git tag mytag. Then at the end of the day, do the original process in reverse but take your diff from the tab: git fp mytag > homeward.patch

And on your home machine, remember to kill the 'work in progress' commit before you apply the patch, with 'git reset --hard HEAD^'. That's a hard reset this time, since the changes in that commit are in the new commits you're bringing back.

Hope your Drupal Coworker Friday was as productive as mine!

Subscribe to RSS - git