Patch and Commit message guidelines -- first draft

Fri Mar 4 17:44:07 UTC 2011

In the last meeting I offered to pull together some guidelines for 'good' patch
and commit messages.  Below is my attempt at doing this.  The guidelines are
from many of Wind River's internal guidelines, as well as simply general
knowledge picked up over time.

I'm sure I've missed a few things below, but I hope this is fairly close to what
is intended.

----

This set of guidelines is intended to help both the developer and reviewers of
changes determine reasonable patch and change commit messages.  Often when
working with the code, you forget that not everyone is as familiar with the
problem and/or fix as you are.  Often the next person in the code doesn't
understand what or why something is done so they quickly look at patch and
commit messages.  Unless these messages are clear it will be difficult to
understand the relevance of a given change and how future changes may impact
previous decisions.

By following these guidelines we will have a better record of the problems and
solutions made over the course of development.  It will also help establish a
clear provenance of all of the code and changes within the development.

General Information
-------------------

While specific to the Linux kernel development, the following could also be
considered a general guide for any Open Source development:

http://ldn.linuxfoundation.org/book/how-participate-linux-community

May of the guidelines in this document are related to the items in that information.

Pay particular attention to section 5.3 that talks about patch preparation.
They key thing to remember is to break up your changes into logical sections.
Otherwise you run the risk of not being able to even explain the purpose of a
change in the patch headers!

Patch and Commit Headers
------------------------
There seems to always be a question or two surrounding what a person
should put in a patch header, or commit message.

The general rules always apply, those being that there is a single line
short log (think of this as the subject when the patch is e-mailed), and then
the more detailed long log, and closure with tag lines like the
"Signed-off-by:".

New development
---------------

A minimal patch or commit message would be of the format:

---
foobar: Adjusted the foo setting in bar

The foo setting in bar was adjusted to enable more foobar resources.  The new
value was determined by determining the amount of free resources and determining
the best value to use up all of system memory.

Signed-off-by: Joe Developer <joe.developer at example.com>
---

The minimal commit is good for new code development and simple changes.

The single short log message should begin with an indicator as to the primary
item changed by this patch, followed by summary of the change.  In the above
case we're indicating that we've changed the "foobar" item, by "adjusting the
foo setting in bar".

Optionally, you may include pointers to defects this change corrects.  In the
above, there was not bug number included.  The format of this reference within
Open Embedded is not yet defined.  [MGH: Is this true?  Is there a better way to
clarify this?]

You must then have a full description of the change.

Finally one or more sign-off-by lines should exist.  Each developer responsible
for working on the patch is responsible for adding a Signed-off-by: line.

It is not acceptable to have an empty or non-existent header, or just a single
line message.  The summary and description is required for all changes.

Importing from elsewhere
-----------------------------
If you are importing work from somewhere else, the minimum commit message is not
enough.  It does not clearly establish the provenance of the code.

If you are pulling in a copy of code from another source, it is required that
you full document the source as part of the commit message.  Such as:

---
foobar: Import foobar from Open Embedded

The foobar recipe was imported from the Open Embedded git server
(git://git.openembedded.org/openembedded) as of commit id
b65a0e0c84cf489bfa00d6aa6c48abc5a237100f.

Signed-off-by: Your Name <your.name at openembedded.org>
---

If the item being imported is a patch or commit from another source, you
similarly have to identify the source as part of your commit message as well as
preserve the upstream commit information.

By default you should keep the original authors summary and commit message,
along with any sign-off information.  You are also required to sign-off on the
change and indicate where this code came from and any changes that you made.

A message similar to the following would be created if you base your changes off
of the commit above:

---
foobar: Adjusted the foo setting in bar

The foo setting in bar was adjusted to enable more foobar resources.  The new
value was determined by determining the amount of free resources and determining
the best value to use up all of system memory.

Signed-off-by: Joe Developer <joe.developer at example.com>

Code was imported from the foobar git (git://example.com/git/foo), commit id
b65a0e0c84cf489bfa00d6aa6c48abc5a237100f.

The settings to bar were adjusted by 2% as necessary to correct the over use of
the core memory.

Signed-off-by: Your Name <your.name at openembedded.org>
---

If the original patch or change that was imported does not have a summary and
commit message from the original author, it is still your responsibility to add
them to the patch.  Just as if you wrote the code should be able to clearly
explain what the change does.  You are still required to clearly document where
the patch originated from.

Common Errors in Patch and Commit messages
------------------------------------------

 - Don't simply translate your C code change into English for a commit
log. The log "Change compare from zero to one" is bad because it
describes only the code change in the patch; we can see that from
reading the patch itself. Let the code tell the story of the mechanics
of the change (as much as possible), and let your comment tell the
other details -- i.e. what the problem was, how it manifested itself
(symptoms), and if necessary, the justification for why the fix was
done in manner that it was.

- Don't start your commit log with "This patch..." -- we already know
it is a patch.

- Don't repeat your short log in the long log. If you really really
don't have anything new and informational to add in as a long log,
then don't put a long log at all. This should be uncommon -- i.e.
the only acceptable cases for no long log would be something like
"Fix spelling mistakes in Documentation/README" or similar.

- Don't put absolute paths to source files that are specific to your site,
i.e. if you get a compile error on "fs/exec.c" then tidy up the parts
of the compile output to just show that portion. We don't need to see
that it was /home/wally/src/git/linux/devel/2009/Nov/linux-2.6/fs/exec.c
or similar - that full path has no value to others.

- Always use the most significant ramification of the change in the
words of your subject/shortlog. For example, don't say "fix compile
warning in foo" when the compiler warning was really telling us that
we were dereferencing complete garbage off in the weeds that could
in theory cause an OOPS under some circumstances. When people are
choosing commits for backports to stable or distro kernels, the
shortlog will be what they use for an initial sorting selection.
If they see "Fix possible OOPS in...." then these people will look
closer, whereas they most likely will skip over simple warning fixes.

- Don't put in the full 20 or more lines of a backtrace when really it is
just the last 5 or so function calls that are relevant to the crash/fix.
If the entry point, or start of the trace is relevant (i.e. a syscall
or similar), you can leave that, and then replace a chunk of the
intermediate calls in the middle with a single [...]

- Don't include timestamps or other unnecessary information, unless they are
relevant to the situation (i.e. indicating an unacceptable delay in
a driver initialization etc.)