This page is primarily intended for Mercurial's developers.
How to help us improve Mercurial's code.
1. Submission checklist
Please double-check your patch meets the following guidelines (explained below) before hitting send:
first line of commit message is of the form "topic: uncapitalized, no trailing period"
bugs that are resolved are mentioned in summary in the form "(issueNNNN)" (no space)
if emailing, patches are sent to the mercurial-devel mailing list (not the BTS) and are properly formatted (see below)
patch does just one thing (if you need a bullet list, split your patch)
- patch contains no other gratuitous changes:
- whitespace changes
- code movement
- typo fixes
- string changes
- patch leaves Mercurial in a working state (doesn't depend on future patches)
test suite has been updated and runs cleanly
code passes check-code and matches coding style (no foo_bar variable names!)
relevant help text is updated (see mercurial/help/)
- appropriate branch is indicated in email subject (eg hg email --flag stable)
if a major release code freeze is in effect (see TimeBasedReleasePlan) patches intended for the default branch should not be sent; instead, send them immediately after the release is completed
- version is indicated in email subject if this is an amended patch, (eg hg email --flag v2 if it has been amended once, or v3, v4, etc.)
2. Getting started
Start by cloning a copy of Mercurial's main development repo:
hg clone https://www.mercurial-scm.org/repo/hg-committed
You'll also want to read the following pages:
Other pages you may find important:
3. Organizing patches
If your first submission is not 'minimal', you will probably be sent back here. Save yourself time and start small!
If you're making a large change, we're probably going to want it broken into a series of smaller patches (see OneChangePerPatch). This makes for easier review and tests both for us and for you. This can be tricky at first and you might find tools like MQ and record useful in this process.
Each patch in a series should:
- implement one clear step in your process
- leave the code in a working state (so bisect always works!)
- include relevant test changes so they're independently testable
- be sent as a separate email
Do not mix formatting changes, organizational changes, or multiple functional changes in the same patch!
Things to consider:
- put the least controversial pieces first - if you're lucky, they'll get applied right away
- the difficulty of reviewing a patch increases rapidly with size - small patches are more likely to get attention
- the probability of a patch getting rejected is also a function of its size - smaller patches mean fewer review round-trips
- bigger patches have more bugs - smaller patches make it easier to locate regressions
3.1. Editing history
You will almost certainly find it necessary to do some form of history editing to generate clean commits, especially after you get your first review feedback. There are multiple tools of varying degrees of power available for this purpose:
commit --amend - modify a head commit's description and contents
rebase - move your commits from one part of history to another, combine commits
mq - manage a stack of changes as patches
histedit - interactively rearrange and modify your commits
evolve - our vision for how history editing should work in the future (experimental)
4. Coding style and testing
If you send a patch with an underscore in a variable name, we'll know you haven't read this page!
See our coding style for what we expect code to look like (yes, we're serious about the underscores)
Use contrib/check-code.py to check for common style errors ( python contrib/check-code.py --blame `/usr/bin/hg manifest` )
Add new test cases as needed
Run the test suite to make sure you haven't broken anything
- Patches should apply cleanly against the tip of the appropriate branch (default or stable)
Don't touch the i18n/ directory for code or doc changes (translations is a separate process)
5. Example patch
Here is an example of a Mercurial patch with proper summary, description, coding style, and testing.
changeset: 21111:dda11e799529 branch: stable user: FUJIWARA Katsunori <firstname.lastname@example.org> date: Tue Mar 25 19:34:17 2014 +0900 files: mercurial/hg.py tests/test-clone.t description: hg: use "os.path.join()" to join path components which may be empty (issue4203) Changset 2d0ab571b822 rewriting "hg.copystore()" with vfs uses 'dstbase + "/lock"' instead of "os.path.join()", because target files given from "store.copyfiles()" already uses "/" as path separator But in the repository using revlog format 0, "dstbase" becomes empty ("data" directory is located under ".hg" directly), and 'dstbase + "/lock"' is treated as "/lock": in almost all cases, write access to "/lock" causes "permission denied". This patch uses "os.path.join()" to join path components which may be empty in "hg.copystore()". diff -r c57c9cece645 -r dda11e799529 mercurial/hg.py --- a/mercurial/hg.py Mon Mar 24 21:27:40 2014 -0400 +++ b/mercurial/hg.py Tue Mar 25 19:34:17 2014 +0900 @@ -213,8 +213,10 @@ dstvfs.mkdir(dstbase) if srcvfs.exists(f): if f.endswith('data'): + # 'dstbase' may be empty (e.g. revlog format 0) + lockfile = os.path.join(dstbase, "lock") # lock to avoid premature writing to the target - destlock = lock.lock(dstvfs, dstbase + "/lock") + destlock = lock.lock(dstvfs, lockfile) hardlink, n = util.copyfiles(srcvfs.join(f), dstvfs.join(f), hardlink) num += n diff -r c57c9cece645 -r dda11e799529 tests/test-clone.t --- a/tests/test-clone.t Mon Mar 24 21:27:40 2014 -0400 +++ b/tests/test-clone.t Tue Mar 25 19:34:17 2014 +0900 @@ -621,3 +621,17 @@ #endif $ cd .. + +Test clone from the repository in (emulated) revlog format 0 (issue4203): + + $ mkdir issue4203 + $ mkdir -p src/.hg + $ echo foo > src/foo + $ hg -R src add src/foo + $ hg -R src commit -m '#0' + $ hg -R src log -q + 0:e1bab28bca43 + $ hg clone -U -q src dst + $ hg -R dst log -q + 0:e1bab28bca43 + $ cd ..
6. Changing C code
Changes involving the C code should be done so that the tests will pass across changesets even without recompiling. This allows "bisect" to be run without needing to run "make" each time, for example.
A recommended strategy for changing behavior in the C code is to add a new interface to the C code when changing behavior (do not change old interfaces). Then in the Python code, check for the existence of the new interface.
If you change tests that rely on changes to C extensions, make sure you also run the tests in "pure" Python mode (and vice versa). For example:
$ python run-tests.py --pure
7. Patch descriptions
It's important that you describe your patch. Patch descriptions should be in the following format:
opener: check hardlink count reporting (issue1866) The Linux CIFS kernel driver (even in 2.6.36) suffers from a hardlink count blindness bug (lstat() returning 1 in st_nlink when it is expected to return >1), which causes repository corruption if Mercurial running ...
- lowercase summary line, no trailing period
- start with the most useful topic keyword (eg command name, subsystem), (the keywords mostly exist to sort our end-user release notes).
- summarize the fix, not the problem
add '(issueNNNN)' if it fixes an issue in the BugTracker (and automatically move the issue to testing)
- use '(BC)' to flag backward compatibility changes, use '(API)' to flag major internal API changes.
- a blank line after the summary
- a more complete description of the problem if necessary
- all lines less than 80 characters
Patch descriptions should be aimed at helping the reviewer understand the issue you're addressing.
Try to use the form "When I tried to do X, I got result Y, but the result should be Z". This is better than "X does not work" which assumes a common understanding of what it means for X to work and leaves the reader to intuit what Y and Z might have been.
Try to answer the following, where appropriate:
- why we need this patch
- how you've implemented it
- why the choices you've made are the right ones
- why the choices you didn't make are the wrong ones
- what all the corner cases are (consider a table!)
- what shortcomings exist
- what file formats and data structures you've used
what compatibility issues exist
- what's missing, if anything
what it looks like, if relevant (include sample output!)
8. Sending patches
Sending a patch implies granting permission to use it in our project under an appropriate license
There are multiple options for sending patches to Mercurial. In order of preference:
If you have already multiple series waiting for review, sending any more patches is not advised. Please, please, please read the Flow control section.
Because this is a community project and our developers are very busy, patches will sometimes fall through the cracks. If you've gotten no response to your patch after a few days, feel free to resend it. Find the patch submission on the email@example.com and reply to it, kindly requesting someone look at it. If it was submitted to Phabricator, you can leave a common on Phabricator.
8.1. Emailing patches
Patches go to firstname.lastname@example.org - no subscription necessary! (we manually whitelist all legitimate posters)
Patch emails should have [PATCH] in the subject followed by a summary (not included in the patch). New versions of the patch should be flagged with V2, V3 etc. so it becomes e.g. [PATCH V2]. See --flag option to hg email.
We prefer patches in the message body so we can review them (no attachments or URLs!)
Patches should be in the '# HG changeset patch' form output by 'hg export' - unified diff with author and full patch description (see also the PatchbombExtension)
Apple Mail, Gmail and possibly other mail clients corrupt patches when sending them in the body, even when sending as plain text. Use the patchbomb extension, instead; or push your changes to the pushgate, which will then send them as patches to the mailing list.
The best way to achieve well-formed patches is to use patchbomb extension which automates the process. Add something like the following to your .hgrc:
[extensions] patchbomb= [patchbomb] confirm=True intro=never [email] to = email@example.com from = Ada Lovelace <firstname.lastname@example.org> method = smtp [smtp] host = smtp.gmail.com port = 587 tls = starttls username = email@example.com
Run hg help config.smtp for more SMTP configuration options and hg help email and hg help -e patchbomb for more info on the email command and the extension that provides it.
Then run the following to do a dryrun test:
$ hg email --test <change1> <change2> ...
'--flag' can be used to flag a patch, e.g. "STABLE", "RFC", "v2"
don't use '--inline', it's a weird MIME thing and not what we want
'--in-reply-to' can be used to specify a message-id to preserve threading (helpful for followup patches).
Please do not send a 0 of X summary message. Those will be deleted in the inbox of code reviewers, and never be read. It's totally fine to discuss the history and purpose of a patch in a patch description. Future archaeologists will thank you.
9. Flow control
Please try not to swamp the list with patches. We have many contributors and few reviewers and we'd like to be sure everyone's patches get looked at.
Please also make sure the state of your patches in Patchwork is updated. This system is strictly secondary to the mailing list, but is helpful to other reviewers for finding patches that still need review.
- Avoid sending more than 10 patches at once. If you have a big stack of patches for review, get them reviewed small groups (around 6 patches) by small group. If your stack is very big you probably want to tell people on the list about it to figure out a way to get it reviewed.
Avoid having too many series in review at the same time. If you already have multiple series in review (even unrelated) sending more unrelated series is not going to help. Apply feedback you received on existing series to close them. Or Do some review yourself to help people reviewing your patches. (see ReviewProcess)
10. Etiquette and advice
- Try to respond quickly to feedback before we forget what your patch is about
- Tell us everything you want us to know about a patch, every time
- If we ask for fixes, don't send a patch to your patch, send a new fixed patch
Consult 'hg log' to cc: the most relevant developers for the code you're working on
- Consider giving input on other people's patches
Discuss your patches on IRC to get faster review and valuable initial feedback
11. See also
Our compatibility rules for new features
The patchbomb extension automates emailing patches
The MQ extension is handy for managing patch series