GitDevelopers

Git opens up a number of new options for contributing to OpenAFS. For the first time, it is easy to review code that is pending addition to the OpenAFS tree. In fact, reviewing code is one of the best ways to ensure that the releases that OpenAFS ships remain stable and functional. If you are interested purely in reviewing, then please skip to that section towards the end of this document.

Git also changes the way that developers interact with the OpenAFS tree. Instead of just having a single version of the tree on your local machine, you have a compressed copy of the entire repository. Additionally, you no longer have to produce patches to send code upstream - any developer can push into the OpenAFS repository directly, through gerrit, our code review tool.

Whilst git is a far more powerful tool than CVS it is also, inevitably, more complex. This document can only scratch the surface of what's possible with git - there are many, many, documents available that describe git in greater detail, and references to some of them are provided at the end.

Getting git

Firstly, if your machine doesn't already have it installed, get a copy of the 'git' version control system. This is available for many platforms from their upstream package repositories or, failing that, can be downloaded in both source and binary form from

Getting the OpenAFS repository

You can download the entire OpenAFS repository by running

git clone git://git.openafs.org/openafs.git

to place the clone in a directory called 'openafs' or

git clone git://git.openafs.org/openafs.git <name-of-directory>

to place your clone in a specific directory

This will give you a complete copy of the OpenAFS repository and unless you are exceptionally short of either disk space, or time, is the approach we recommend. Unlike CVS, you are not just checking out a particular branch, but making a local copy of the project's whole revision history, across all of it's branches. This does make the download pretty large - around 150Mbytes at the time of writing.

Updating the local copy

When you want to update the local repository with the central OpenAFS one, running

git pull

will pull all of the new changes into your local repository, and merge those changes into your current working tree. Note that whilst this is fine when you are browsing the repository, you may want to exercise more control over how upstream changes are merged into your development code.

Checkout a particular branch

The OpenAFS repository contains many branches, most of which are historical and should not be used for new development. Current development should be targetted at 'master' (for feature work, and general bugfixing), or at 'openafs-stable-1_4_x' (bug fixes specific to the current stable release). Note that the openafs-devel-1_5_x branch is now effectively dead - future 1.5 releases will occur from the 'master' branch.

A complete list of all branches in the upstream OpenAFS repository can be obtained by running

git branch -r

If all you wish to do is browse code, then you can directly check out these remote branches, using

git checkout origin/<branch>

For example, to checkout the 'openafs-stable-1_4_x' branch:

git checkout origin/openafs-stable-1_4_x

Note that if you wish to do development on a particular branch, you should either make a local branch which tracks the remote one, using something like

git checkout -b openafs-stable-1_4_x origin/openafs-stable-1_4_x

or, more simply

git checkout --track origin/openafs-stable-1_4_x

or by creating a topic branch as discussed below.

Checkout a particular release

Every release version of ?OpenAFS is marked in the repository by means of a tag.

A complete list of all tags can be obtained by running

git tag

To checkout a particular tag

git checkout openafs-stable-1_4_10

Again, whilst a direct checkout of a remote tag is fine for code browsing, it should not be used as a place to start development. If you must do development against a tag, then create a local topic branch with it as a starting point, as is discussed below. However, in general, please don't develop from a particular tag, but instead work from a branch tip. It makes it much easier to integrate your changes!

Viewing deltas

OpenAFS's original CVS repository used the concept of deltas as a means of grouping a large number of related changes into a single item, which could be easily fetched and referred to. In git, a delta should be simply a single commit. Deltas are represented by means of a special form of git tag, allowing you to locally view the change and commit message that corresponds to each one. In order to keep down the transfer size, deltas are not included in the repository you get when you do a git clone - there are over 10,000 delta references, and having them in your local repository can cause performance issues. If you really wish to be able to locally browse deltas, then run the following

git config --add remote.origin.fetch '+refs/deltas/*:refs/remotes/deltas/*'
git fetch origin

You can then view a specific delta by doing

git show refs/remotes/deltas/<branch>/<delta>

Sadly, historical accidents mean that not all of our deltas can be represented by means of single commit. Where this is the case, a delta-name will have a trailing -part-, where each of these numbers must be used to form the complete delta. This only applies to some deltas created before the git conversion - all deltas created from now on will be single commits.

Introducing yourself to git

Before you begin development, you should let git know who you are. This provides it with a name, and email address, that is used to attribute all commits in your repository, and in any that you share code with.

git config user.name "Joe Bloggs"
git config user.email "joe.bloggs@example.org"

If you want to make this settings for all of your repositories, then add the --global switch.

git config --global user.name "Joe Bloggs"
git config --global user.email "joe.bloggs@example.org"

Note that this email address is the address by which you will be identified in ?OpenAFS's revision history - it is also the address to which the gerrit code review tool will send all email related to the review of your code.

If you plan on making changes to OpenAFS (and why else would you be reading this?) you should probably also grab The change id hook described in Registering With gerrit below. You can grab and apply the hook before registering, and it'll make sure your pre-registration development has the appropriate change IDs in the log. The hook only applies to your openafs development, so you're not going to mess up any of your non-OpenAFS work.

Helpful git tips

Here are a few other git settings that may be helpful when working with the source.

Prevent C labels from being treated as function names by git diff:

git config diff.default.xfuncname '^[[:alpha:]$_].*[^:]$'

Changes the style used to indicate merge conflicts in source files:

git config merge.conflictstyle diff3

Whitespace handling settings:

git config apply.whitespace fix
git config core.whitespace trailing-space,space-before-tab,indent-with-non-tab
git config config.cleanup whitespace

Starting development

We strongly recommend that you do all of your development upon 'topic branches' This allows you to isolate multiple unrelated changes, and makes it easier to keep your tree in sync with the upstream ?OpenAFS one.

Before creating a new topic branch, running

git fetch

will make sure that your repository knows about the latest upstream changes (unlike git pull, this will not update any files that you may have checked out)

To create a new topic branch:

git checkout -b <branch>

For example, to work on a patch to fix printf warnings, based on the current development code, I would do:

git checkout -b fix-printf-warnings origin/master

This puts me on a new branch, ready to start writing code. All new development should be based upon the origin/master branch, submissions based upon other branches are unlikely to be accepted, unless they address issues that are solely present in that branch.

'git add' is used to tell git about any new files you create as part of your patch. If your patch results in any new compilation products (object files, new executables, etc) that git should not be tracking, please make sure that they're caught by the .gitignore mechanism. You can do this by checking that they don't appear in the output from 'git status'.

'git mv' and 'git rm' are used to move and delete files respectively.

'git commit -a' is used to commit code to all of the files that git is currently tracking (that is, all of the files that you have checked out from the repository, and all those which you have run git add on)

When you can't see the wood for the trees

If, in the middle of development, you discover that you've gone down a blind alley, and wish to go back to the state of your last commit

git reset --hard

will discard all of the changes you have made since the last commit, or

git checkout -f <file>

will restore <file> to the state it was in at the last commit.

Keeping up with the Joneses

If you're working on a long running development project, you will find that the point your created your topic branch rapidly recedes into history. At some point (and at least before you share your code with us), you'll probably want to update your tree. There are a number of ways of doing this.

If you haven't shared your tree with anyone else, then you can use

git rebase <branch> <topic>

(Where <branch> is the name of the upstream branch - for example origin/master, and <topic> is the name of the topic branch you are currently working on)

Note that git rebase changes your local history (it moves the branch point of your topic branch to the tip of the upstream branch), and is a bad idea if others have cloned your repository. See 'man git-rebase' for more details.

If you can't rebase, then consider either merging the changes onto your local branch, or creating a new topic branch and cherry picking your changes onto it. The man pages for 'git merge' and 'git cherry-pick' provide more detail on these options.

Sharing your code with us

How you work from this point onwards depends on how you intend making your code available to OpenAFS. We're still happy to receive submission by patch (by sending your changes to openafs-bugs@openafs.org), but it makes it much easier for us if you push directly from your git tree to our code review system, gerrit.

Registering with gerrit

To register with gerrit, visit http://gerrit.openafs.org/ and log in using any OpenID. OpenIDs are available from a large number of internet services, including Google Accounts. Note that we're aware of problems using LiveJournal OpenIDs to access the service.

If, when you introduced yourself to git, you gave it a different email address than the one gerrit currently has listed for you, you need to introduce yourself to gerrit under that name, too. Click on 'Contact Information', and select "Register New Email...". If gerrit thinks you are an "Anonymous Coward" then you can fix that on this page as well.

In order to be able to upload code, you now need to create a ssh key that gerrit can use to identify you, or tell gerrit about one that already exists.

To create a new ssh key, if you don't already have one, run

ssh-keygen -t rsa -f ~/.ssh/id_rsa

The public key for this is now stored in ~/.ssh/id_rsa.pub.

To tell gerrit about your key, log in, and go to 'Settings'. Select 'SSH Keys', and paste your public key into the "Add SSH Public Key" box. Click on 'Add' to add the new public key. Now go to 'Profile' and make a note of the Username that is listed on that page.

To make things easier, set up OpenSSH so that it knows about the defaults for the gerrit server. Edit ~/.ssh/config, and add a section like:

Host gerrit.openafs.org
User <Username>
IdentityFile ~/.ssh/id_rsa
Port 29418
HostKeyAlgorithms +ssh-rsa
PubkeyAcceptedAlgorithms +ssh-rsa

(where Username is the username you noted down from the 'Profile' page)

The change id hook

Gerrit introduces the concept of "change IDs". This is a unique reference for a particular change, which remains constant regardless of any changes that are made to the implementation. This allows a single reference to be attached to a given modification, irrespective of any rewrites that may occur as a result of review comments. Manually maintaining change Ids is a pain, so gerrit provides a git hook which can be used to automatically add a change Id to any new modifications you create.

The hook should be downloaded from the ?OpenAFS gerrit server by running the following, in the top level of your git tree

scp -O -p -P 29418 gerrit.openafs.org:hooks/commit-msg .git/hooks/

Uploading to gerrit

When submitting to gerrit, it's important to realise that each commit in your branch will become a changeset in the upstream OpenAFS, typically with no modification at our end. It is therefore important that these commits follow some simple rules...

First, each commit should be complete. The maxim "one change per commit, one commit per change" applies here. Each commit should build and test in its own right. Typically, this means that a change will be in a small number of commits. If, during development, you have created many more than this (for example, you've created a large number of bug fix commits), please use 'git rebase', or cherry pick these commits into a separate tree before uploading them. Note, however, that "one change" could equate to a change to source code and a change to the corresponding documentation for that code specific change.

Secondly, each commit should have a meaningful revision log. The internals of git means that we can't easily edit these before pushing them into the tree, so we'd like you to get them right for us! A commit message should be comprised of a single 'subject' line (which must not end with a full stop), followed by a blank line, followed by one or more paragraphs explaining the purpose of the patch. Gerrit warns if the 'subject' is longer than 65 characters. If it is intended to fix a bug in OpenAFS RT, then the word 'FIXES' followed by the bug number or comma-separated list of bug numbers should be included on a line of its own. The 'LICENSE' keyword can be used to indicate code which is covered under a license other than the IPL, although please speak to a gatekeeper if you intend using this. An example commit message would be

  Add option to disable syscall probing

  Add a --disable-linux-syscall-probing flag to allow the probing of
  kernel memory for the syscall table to be disabled at compile time. This
  helps with Linux architectures which have compile, run or load time
  issues with syscall probing by providing a band-aid fix until we can
  improve the configuration logic in this area.

  FIXES 123456

Thirdly, each commit should have a valid changeID. Manually maintaining these is difficult and error prone, so we would strong advise that you install the changeID hook detailed earlier. This will automatically add a ?ChangeId line to your commit message if it doesn't already contain one.

Fourthly, each commit must adhere to the OpenAFS whitespace policy whereby new commits will not be accepted if they have trailing spaces, spaces before tabs, or indentation without tabs for new files and functions. The prevailing formating conventions for exsiting code should be followed. Unrelated formatting changes to code is discouraged.

Git can be configured to highlight the whitespace policy violation with the following global setting:

git config --global core.whitespace trailing-space,space-before-tab,indent-with-non-tab

and

git rebase --whitespace=fix

can be used to automatically fix any policy violations on your local branch before pushing the changes to Gerrit. Finally, Git 1.8.2 and above can be configured to apply this policy to all local commits:

git config --global config.cleanup whitespace

Then commit your changes, for example with a file that contains your log message

git commit -a -F <your-log-message-in-this-file>

(where <your-log-message-in-this-file> is a file with the log message)

Once your commits have a proper commit message and have all whitespace errors fixed, use

git log -p origin/<branch>..HEAD

(where <branch> is the upstream branch upon which you are basing this patch).

to check that what you're giving us makes sense. Then, upload the commits to gerrit using

git push ssh://gerrit.openafs.org/openafs.git HEAD:refs/for/<branch>/<topic>

(again <branch> is the name of the upstream branch that you are pushing the changes into, not the name of any local branch you may have been developing on and <topic> is an optional name that can be used to group your commits together for easier reviewing.)

In this case, refs/for is a literal string. So, if you had been developing against the "master" branch and the change replaced "strcpy" with "strlcpy", you might upload your changes with:

git push ssh://gerrit.openafs.org/openafs.git HEAD:refs/for/master/strcpy-to-strlcpy

Although, it would be sufficient to simply issue the command as:

git push ssh://gerrit.openafs.org/openafs.git HEAD:refs/for/master

This relies upon the ssh configuration you performed earlier. If it fails to work, please consult the troubleshooting notes at http://gerrit.googlecode.com/svn/documentation/2.0/user-upload.html

Assuming all has gone well, this will have added the entry to the code review queue. The output from git review will give you a change number - this is a unique reference for this particular set of changes. During review you'll be emailed with any comments anyone makes, and can respond to those comments using the gerrit web interface (see the section on reviewing, below). It's possible that issues with your change may be noticed during the review process, and you may be asked to revise it, or update changes to the tip of the tree.

Revising your change

It's possible that your modifications won't be accepted first time. In this case, you need to revise your changes, and resubmit them to gerrit. Please note that this should always be done by modifying your original changeset, not by submitting a new change that makes the required fixes. Either git commit -a --amend, or git rebase should be used to combine your changes with the original changeset, and then you should push this to gerrit with

git push ssh://gerrit.openafs.org/openafs.git <revision>:refs/for/<branch>

where <revision> is the SHA-1 hash of the revised change that follows the word commit in the log message, or simply HEAD if the revised change is the most recent change on your topic branch. You can obtain the SHA-1 hash of a commit by using 'git log'. Note that pushing to refs/for/... requires a change-id in your commit message, so that Gerrit can match to the new change with the one you submitted previously. See http://gerrit.googlecode.com/svn/documentation/2.0/user-upload.html for full details.

Updating your change

It's possible that your change may have been made against a tree which is too old for it to apply to the tip. In this case, gerrit will let you know that there is a collision, and request that you update the change to the tip.

You can do this with

git rebase origin/master <topic>

(assuming your patch is against the 'master' git branch, and lives on the <topic> branch)

When a rebase is performed there may be conflicts that cannot be automatically resolved by git. The default style of conflict resolution displays the current version of the code on HEAD and the version from the commit that is being rebased. This level of detail is often insufficient to determine how to resolve the conflict. Switching to conflict style "diff3" will also show the original version of the code which your commit modified. Turn on "diff3" by applying the following configuration setting:

git config --global merge.conflictstyle diff3

After you have resolved all conflicts and are once again happy with the commit, simply resubmit your change in the same way as if you had been asked to revise it (see notes above)

Pulling Up a Change to a Stable Branch

After a change has passed through the Gerrit Code Review process it will be merged onto the master branch of the OpenAFS repository. The merged commit will then become visible in your local repository the next time a

git pull

command is executed. OpenAFS releases are not issued from the master branch. Instead releases such as 1.6.11 are issued from stable branches such as openafs-stable-1_6_x. After a change is approved on master the change must be pulled up to the stable branch, be modified for any differences between master and openafs-stable-N_N_x and be resubmitted to Gerrit.

Start by checking out the stable branch of interest. For example, if you want to submit a change to the next 1.6 series release you would first checkout master.

git checkout --track origin/master

or just

git checkout master

if you have already created the tracking branch. Then

git pull

to import any new changes to the local repository and fast-forward HEAD to the most recent commit merged onto origin/master. It is important to note that if any local changes were committed onto the local master tracking branch that a fast-forward operation will not occur. Instead the local changes and the changes to origin will be merged producing a merge commit. A merge commit is not compatible with the OpenAFS management of the git repository. To rebase local changes onto of the HEAD of origin/master the command

git pull --rebase

can be performed. If there are conflicts when applying the local changes to current HEAD of origin/master you will have an opportunity to resolve them.

Once the change you wish to pullup to the stable branch is in the local repository you can use

git log master

to find it and record the sha1 identifier for the commit. Now you are ready to switch to the stable branch for example, openafs-stable-1_6_x. If you do not already have a local tracking branch execute

git checkout --track origin/<stable-branch>

and if you do then execute

git checkout <stable-branch>

After switching branches you might be told that the local branch is behind the origin. If so, execute

git pull

to fast-forward the branch HEAD to the most recent commit. Now you can pull the commit from master by using performing a cherry-pick:

git cherry-pick -x <sha1>

This will create a new commit on the current branch with a commit message that has been modified with

(cherry picked from commit <sha1>)

This commit message will include the Change-Id from the original submission to Gerrit. It is important that the Change-Id values submitted to Gerrit be unique but Gerrit will not enforce this. You must remember to edit the commit message and remove the old Change-Id using

git commit --amend

so that the Change ID Hook will generate a new Change-Id before pushing the change to Gerrit.

Once the commit message has been ammended and any other required changes are made it is time to push the change to Gerrit:

git push ssh://gerrit.openafs.org/openafs.git HEAD:refs/for/<stable-branch>/<topic>

The change will then be evaluated by the OpenAFS Release team for incorporation into a future stable release.

Submitting by patch

If all of this seems too daunting (and please don't let it put you off) you can still, of course, submit patches by email.

git diff HEAD

will give you the set of changes if you don't do local commits. If you make topic branches, and commit things locally, but don't want to go through the whole gerrit process,

git diff master..<topic>

will give all of the changes between the branch point of you topic branch (assuming you branched from 'master') and the last commit.

A better approach is to generate a patch file. To do so commit your changes to a local branch in your repository as you would if you were submitting to gerrit. If your changes are against the "master" branch, instead of pushing the patch execute the command:

git format-patch origin/master..HEAD

For each commit on your local branch after the most recent patch on "master", a separate patch file will be generated.

Regardless of which approach you use, you can e-mail the changes to openafs-bugs@openafs.org as before. Note, however, by doing this you're making someone else take the patch, create a topic branch in their local tree, apply the patch, push it into gerrit, and become responsible for managing the review process. Things would be much more efficient if you pushed into gerrit yourself. Please?

Reviewing changes

We'll now look at how changes that have made it into gerrit can be reviewed. All code review now happens via the http://gerrit.openafs.org interface. You should log in there as detailed above (using any OpenID), and make sure that the email address points to somewhere you'll read regularly.

You'll be presented with a list of patches requiring review or, if someone has asked, patches you've been explicitly requested to review. There are two types of review - Code Review and Verification. Code Review means that you have read through the code, and are satisfied that it works properly, follows the tree's style, and generally doesn't suck. Verification means that you have taken a copy of the patch and tested it. We hope to eventually automate the verification step, but for now both must be performed by hand.

To perform a code review, go through each of the diffs in the current changeset for the code you have decided to review. You can double click on a line to leave a comment. Once you have completed commenting, click on the 'Review' button that's about 3/4 of the way down the page containing the list of patch sets. You will then be asked to score the patch, with a range from -1 to +1. -1 means that you don't think the code should be applied, +1 means that it is good to apply. You can also leave further, general, comments for the patch submitter.

Note that no matter how many +1 or -1 comments a patch receives, the gatekeepers can override these to either permit or forbid submission. Also, at least one gatekeeper must approve a patch before it can be submitted to the tree.

To verify the code, pull the git copy into your local tree, using the git command listed as part of the patch details. For sanity's sake, we'd strongly recommend you do this pull into a topic branch you've created for the purpose. Build it, test it, and report your results. Note that a single -1 to verification can block patch submission, so please use these options wisely. If in doubt, score 0 and leave your comments. Please indicate when verifying which platforms you have tested on, and what testing you performed.

And that's pretty much it. All of this is very new. If you encounter any problems at all, please ask for help on IRC in #openafs, Jabber in openafs@conference.openafs.org or on openafs-devel@openafs.org. Private pleas may be addressed to simon@sxw.org.uk

Acknowledgments

Thanks to everyone who has reviewed this document, and offered corrections and contributions.

-- Simon Wilkinson - 07 Jul 2009