Chapter 12
Managing change with Mercurial Queues

12.1 The patch management problem

Here is a common scenario: you need to install a software package from source, but you find a bug that you must fix in the source before you can start using the package. You make your changes, forget about the package for a while, and a few months later you need to upgrade to a newer version of the package. If the newer version of the package still has the bug, you must extract your fix from the older source tree and apply it against the newer version. This is a tedious task, and it’s easy to make mistakes.

This is a simple case of the “patch management” problem. You have an “upstream” source tree that you can’t change; you need to make some local changes on top of the upstream tree; and you’d like to be able to keep those changes separate, so that you can apply them to newer versions of the upstream source.

The patch management problem arises in many situations. Probably the most visible is that a user of an open source software project will contribute a bug fix or new feature to the project’s maintainers in the form of a patch.

Distributors of operating systems that include open source software often need to make changes to the packages they distribute so that they will build properly in their environments.

When you have few changes to maintain, it is easy to manage a single patch using the standard diff and patch programs (see section 12.4 for a discussion of these tools). Once the number of changes grows, it starts to makes sense to maintain patches as discrete “chunks of work,” so that for example a single patch will contain only one bug fix (the patch might modify several files, but it’s doing “only one thing”), and you may have a number of such patches for different bugs you need fixed and local changes you require. In this situation, if you submit a bug fix patch to the upstream maintainers of a package and they include your fix in a subsequent release, you can simply drop that single patch when you’re updating to the newer release.

Maintaining a single patch against an upstream tree is a little tedious and error-prone, but not difficult. However, the complexity of the problem grows rapidly as the number of patches you have to maintain increases. With more than a tiny number of patches in hand, understanding which ones you have applied and maintaining them moves from messy to overwhelming.

Fortunately, Mercurial includes a powerful extension, Mercurial Queues (or simply “MQ”), that massively simplifies the patch management problem.

12.2 The prehistory of Mercurial Queues

During the late 1990s, several Linux kernel developers started to maintain “patch series” that modified the behaviour of the Linux kernel. Some of these series were focused on stability, some on feature coverage, and others were more speculative.

The sizes of these patch series grew rapidly. In 2002, Andrew Morton published some shell scripts he had been using to automate the task of managing his patch queues. Andrew was successfully using these scripts to manage hundreds (sometimes thousands) of patches on top of the Linux kernel.

12.2.1 A patchwork quilt

In early 2003, Andreas Gruenbacher and Martin Quinson borrowed the approach of Andrew’s scripts and published a tool called “patchwork quilt” [AG], or simply “quilt” (see [Gru05] for a paper describing it). Because quilt substantially automated patch management, it rapidly gained a large following among open source software developers.

Quilt manages a stack of patches on top of a directory tree. To begin, you tell quilt to manage a directory tree, and tell it which files you want to manage; it stores away the names and contents of those files. To fix a bug, you create a new patch (using a single command), edit the files you need to fix, then “refresh” the patch.

The refresh step causes quilt to scan the directory tree; it updates the patch with all of the changes you have made. You can create another patch on top of the first, which will track the changes required to modify the tree from “tree with one patch applied” to “tree with two patches applied”.

You can change which patches are applied to the tree. If you “pop” a patch, the changes made by that patch will vanish from the directory tree. Quilt remembers which patches you have popped, though, so you can “push” a popped patch again, and the directory tree will be restored to contain the modifications in the patch. Most importantly, you can run the “refresh” command at any time, and the topmost applied patch will be updated. This means that you can, at any time, change both which patches are applied and what modifications those patches make.

Quilt knows nothing about revision control tools, so it works equally well on top of an unpacked tarball or a Subversion repository.

12.2.2 From patchwork quilt to Mercurial Queues

In mid-2005, Chris Mason took the features of quilt and wrote an extension that he called Mercurial Queues, which added quilt-like behaviour to Mercurial.

The key difference between quilt and MQ is that quilt knows nothing about revision control systems, while MQ is integrated into Mercurial. Each patch that you push is represented as a Mercurial changeset. Pop a patch, and the changeset goes away.

Because quilt does not care about revision control tools, it is still a tremendously useful piece of software to know about for situations where you cannot use Mercurial and MQ.

12.3 The huge advantage of MQ

I cannot overstate the value that MQ offers through the unification of patches and revision control.

A major reason that patches have persisted in the free software and open source world—in spite of the availability of increasingly capable revision control tools over the years—is the agility they offer.

Traditional revision control tools make a permanent, irreversible record of everything that you do. While this has great value, it’s also somewhat stifling. If you want to perform a wild-eyed experiment, you have to be careful in how you go about it, or you risk leaving unneeded—or worse, misleading or destabilising—traces of your missteps and errors in the permanent revision record.

By contrast, MQ’s marriage of distributed revision control with patches makes it much easier to isolate your work. Your patches live on top of normal revision history, and you can make them disappear or reappear at will. If you don’t like a patch, you can drop it. If a patch isn’t quite as you want it to be, simply fix it—as many times as you need to, until you have refined it into the form you desire.

As an example, the integration of patches with revision control makes understanding patches and debugging their effects—and their interplay with the code they’re based on—enormously easier. Since every applied patch has an associated changeset, you can use “hg log filename” to see which changesets and patches affected a file. You can use the bisect extension to binary-search through all changesets and applied patches to see where a bug got introduced or fixed. You can use the “hg annotate” command to see which changeset or patch modified a particular line of a source file. And so on.

12.4 Understanding patches

Because MQ doesn’t hide its patch-oriented nature, it is helpful to understand what patches are, and a little about the tools that work with them.

The traditional Unix diff command compares two files, and prints a list of differences between them. The patch command understands these differences as modifications to make to a file. Take a look at figure 12.1 for a simple example of these commands in action.

The type of file that diff generates (and patch takes as input) is called a “patch” or a “diff”; there is no difference between a patch and a diff. (We’ll use the term “patch”, since it’s more commonly used.)

A patch file can start with arbitrary text; the patch command ignores this text, but MQ uses it as the commit message when creating changesets. To find the beginning of the patch content, patch searches for the first line that starts with the string “diff -”.

MQ works with unified diffs (patch can accept several other diff formats, but MQ doesn’t). A unified diff contains two kinds of header. The file header describes the file being modified; it contains the name of the file to modify. When patch sees a new file header, it looks for a file with that name to start modifying.

After the file header comes a series of hunks. Each hunk starts with a header; this identifies the range of line numbers within the file that the hunk should modify. Following the header, a hunk starts and ends with a few (usually three) lines of text from the unmodified file; these are called the context for the hunk. If there’s only a small amount of context between successive hunks, diff doesn’t print a new hunk header; it just runs the hunks together, with a few lines of context between modifications.

Each line of context begins with a space character. Within the hunk, a line that begins with “-” means “remove this line,” while a line that begins with “+” means “insert this line.” For example, a line that is modified is represented by one deletion and one insertion.

We will return to some of the more subtle aspects of patches later (in section 12.6), but you should have enough information now to use MQ.

12.5 Getting started with Mercurial Queues

Because MQ is implemented as an extension, you must explicitly enable before you can use it. (You don’t need to download anything; MQ ships with the standard Mercurial distribution.) To enable MQ, edit your ~/.hgrc file, and add the lines in figure 12.5.

Once the extension is enabled, it will make a number of new commands available. To verify that the extension is working, you can use “hg help” to see if the “hg qinit” command is now available; see the example in figure 12.3.

You can use MQ with any Mercurial repository, and its commands only operate within that repository. To get started, simply prepare the repository using the “hg qinit” command (see figure 12.4). This command creates an empty directory called .hg/patches, where MQ will keep its metadata. As with many Mercurial commands, the “hg qinit” command prints nothing if it succeeds.

12.5.1 Creating a new patch

To begin work on a new patch, use the “hg qnew” command. This command takes one argument, the name of the patch to create. MQ will use this as the name of an actual file in the .hg/patches directory, as you can see in figure 12.5.

Also newly present in the .hg/patches directory are two other files, series and status. The series file lists all of the patches that MQ knows about for this repository, with one patch per line. Mercurial uses the status file for internal book-keeping; it tracks all of the patches that MQ has applied in this repository.

Once you have created your new patch, you can edit files in the working directory as you usually would. All of the normal Mercurial commands, such as “hg diff” and “hg annotate”, work exactly as they did before.

12.5.2 Refreshing a patch

When you reach a point where you want to save your work, use the “hg qrefresh” command (figure 12.5) to update the patch you are working on. This command folds the changes you have made in the working directory into your patch, and updates its corresponding changeset to contain those changes.

You can run “hg qrefresh” as often as you like, so it’s a good way to “checkpoint” your work. Refresh your patch at an opportune time; try an experiment; and if the experiment doesn’t work out, “hg revert” your modifications back to the last time you refreshed.

12.5.3 Stacking and tracking patches

Once you have finished working on a patch, or need to work on another, you can use the “hg qnew” command again to create a new patch. Mercurial will apply this patch on top of your existing patch. See figure 12.8 for an example. Notice that the patch contains the changes in our prior patch as part of its context (you can see this more clearly in the output of “hg annotate”).

So far, with the exception of “hg qnew” and “hg qrefresh”, we’ve been careful to only use regular Mercurial commands. However, MQ provides many commands that are easier to use when you are thinking about patches, as illustrated in figure 12.9:

• The “hg qseries” command lists every patch that MQ knows about in this repository, from oldest to newest (most recently created).
• The “hg qapplied” command lists every patch that MQ has applied in this repository, again from oldest to newest (most recently applied).

12.5.4 Manipulating the patch stack

The previous discussion implied that there must be a difference between “known” and “applied” patches, and there is. MQ can manage a patch without it being applied in the repository.

An applied patch has a corresponding changeset in the repository, and the effects of the patch and changeset are visible in the working directory. You can undo the application of a patch using the “hg qpop” command. MQ still knows about, or manages, a popped patch, but the patch no longer has a corresponding changeset in the repository, and the working directory does not contain the changes made by the patch. Figure 12.10 illustrates the difference between applied and tracked patches.

You can reapply an unapplied, or popped, patch using the “hg qpush” command. This creates a new changeset to correspond to the patch, and the patch’s changes once again become present in the working directory. See figure 12.11 for examples of “hg qpop” and “hg qpush” in action. Notice that once we have popped a patch or two patches, the output of “hg qseries” remains the same, while that of “hg qapplied” has changed.

12.5.5 Pushing and popping many patches

While “hg qpush” and “hg qpop” each operate on a single patch at a time by default, you can push and pop many patches in one go. The -a option to “hg qpush” causes it to push all unapplied patches, while the -a option to “hg qpop” causes it to pop all applied patches. (For some more ways to push and pop many patches, see section 12.7 below.)

12.5.6 Safety checks, and overriding them

Several MQ commands check the working directory before they do anything, and fail if they find any modifications. They do this to ensure that you won’t lose any changes that you have made, but not yet incorporated into a patch. Figure 12.13 illustrates this; the “hg qnew” command will not create a new patch if there are outstanding changes, caused in this case by the “hg add” of file3.

Commands that check the working directory all take an “I know what I’m doing” option, which is always named -f. The exact meaning of -f depends on the command. For example, “hg qnew -f” will incorporate any outstanding changes into the new patch it creates, but “hg qpop -f” will revert modifications to any files affected by the patch that it is popping. Be sure to read the documentation for a command’s -f option before you use it!

12.5.7 Working on several patches at once

The “hg qrefresh” command always refreshes the topmost applied patch. This means that you can suspend work on one patch (by refreshing it), pop or push to make a different patch the top, and work on that patch for a while.

Here’s an example that illustrates how you can use this ability. Let’s say you’re developing a new feature as two patches. The first is a change to the core of your software, and the second—layered on top of the first—changes the user interface to use the code you just added to the core. If you notice a bug in the core while you’re working on the UI patch, it’s easy to fix the core. Simply “hg qrefresh” the UI patch to save your in-progress changes, and “hg qpop” down to the core patch. Fix the core bug, “hg qrefresh” the core patch, and “hg qpush” back to the UI patch to continue where you left off.

12.6 More about patches

MQ uses the GNU patch command to apply patches, so it’s helpful to know a few more detailed aspects of how patch works, and about patches themselves.

12.6.1 The strip count

If you look at the file headers in a patch, you will notice that the pathnames usually have an extra component on the front that isn’t present in the actual path name. This is a holdover from the way that people used to generate patches (people still do this, but it’s somewhat rare with modern revision control tools).

Alice would unpack a tarball, edit her files, then decide that she wanted to create a patch. So she’d rename her working directory, unpack the tarball again (hence the need for the rename), and use the -r and -N options to diff to recursively generate a patch between the unmodified directory and the modified one. The result would be that the name of the unmodified directory would be at the front of the left-hand path in every file header, and the name of the modified directory would be at the front of the right-hand path.

Since someone receiving a patch from the Alices of the net would be unlikely to have unmodified and modified directories with exactly the same names, the patch command has a -p option that indicates the number of leading path name components to strip when trying to apply a patch. This number is called the strip count.

An option of “-p1” means “use a strip count of one”. If patch sees a file name foo/bar/baz in a file header, it will strip foo and try to patch a file named bar/baz. (Strictly speaking, the strip count refers to the number of path separators (and the components that go with them ) to strip. A strip count of one will turn foo/bar into bar, but /foo/bar (notice the extra leading slash) into foo/bar.)

The “standard” strip count for patches is one; almost all patches contain one leading path name component that needs to be stripped. Mercurial’s “hg diff” command generates path names in this form, and the “hg import” command and MQ expect patches to have a strip count of one.

If you receive a patch from someone that you want to add to your patch queue, and the patch needs a strip count other than one, you cannot just “hg qimport” the patch, because “hg qimport” does not yet have a -p option (see Mercurial bug no. 311). Your best bet is to “hg qnew” a patch of your own, then use “patch -pN” to apply their patch, followed by “hg addremove” to pick up any files added or removed by the patch, followed by “hg qrefresh”. This complexity may become unnecessary; see Mercurial bug no. 311 for details.

12.6.2 Strategies for applying a patch

When patch applies a hunk, it tries a handful of successively less accurate strategies to try to make the hunk apply. This falling-back technique often makes it possible to take a patch that was generated against an old version of a file, and apply it against a newer version of that file.

First, patch tries an exact match, where the line numbers, the context, and the text to be modified must apply exactly. If it cannot make an exact match, it tries to find an exact match for the context, without honouring the line numbering information. If this succeeds, it prints a line of output saying that the hunk was applied, but at some offset from the original line number.

If a context-only match fails, patch removes the first and last lines of the context, and tries a reduced context-only match. If the hunk with reduced context succeeds, it prints a message saying that it applied the hunk with a fuzz factor (the number after the fuzz factor indicates how many lines of context patch had to trim before the patch applied).

When neither of these techniques works, patch prints a message saying that the hunk in question was rejected. It saves rejected hunks (also simply called “rejects”) to a file with the same name, and an added .rej extension. It also saves an unmodified copy of the file with a .orig extension; the copy of the file without any extensions will contain any changes made by hunks that did apply cleanly. If you have a patch that modifies foo with six hunks, and one of them fails to apply, you will have: an unmodified foo.orig, a foo.rej containing one hunk, and foo, containing the changes made by the five successful five hunks.

12.6.3 Some quirks of patch representation

There are a few useful things to know about how patch works with files.

• This should already be obvious, but patch cannot handle binary files.
• Neither does it care about the executable bit; it creates new files as readable, but not executable.
• patch treats the removal of a file as a diff between the file to be removed and the empty file. So your idea of “I deleted this file” looks like “every line of this file was deleted” in a patch.
• It treats the addition of a file as a diff between the empty file and the file to be added. So in a patch, your idea of “I added this file” looks like “every line of this file was added”.
• It treats a renamed file as the removal of the old name, and the addition of the new name. This means that renamed files have a big footprint in patches. (Note also that Mercurial does not currently try to infer when files have been renamed or copied in a patch.)
• patch cannot represent empty files, so you cannot use a patch to represent the notion “I added this empty file to the tree”.

12.6.4 Beware the fuzz

While applying a hunk at an offset, or with a fuzz factor, will often be completely successful, these inexact techniques naturally leave open the possibility of corrupting the patched file. The most common cases typically involve applying a patch twice, or at an incorrect location in the file. If patch or “hg qpush” ever mentions an offset or fuzz factor, you should make sure that the modified files are correct afterwards.

It’s often a good idea to refresh a patch that has applied with an offset or fuzz factor; refreshing the patch generates new context information that will make it apply cleanly. I say “often,” not “always,” because sometimes refreshing a patch will make it fail to apply against a different revision of the underlying files. In some cases, such as when you’re maintaining a patch that must sit on top of multiple versions of a source tree, it’s acceptable to have a patch apply with some fuzz, provided you’ve verified the results of the patching process in such cases.

12.6.5 Handling rejection

If “hg qpush” fails to apply a patch, it will print an error message and exit. If it has left .rej files behind, it is usually best to fix up the rejected hunks before you push more patches or do any further work.

If your patch used to apply cleanly, and no longer does because you’ve changed the underlying code that your patches are based on, Mercurial Queues can help; see section 12.8 for details.

Unfortunately, there aren’t any great techniques for dealing with rejected hunks. Most often, you’ll need to view the .rej file and edit the target file, applying the rejected hunks by hand.

If you’re feeling adventurous, Neil Brown, a Linux kernel hacker, wrote a tool called wiggle [Bro], which is more vigorous than patch in its attempts to make a patch apply.

Another Linux kernel hacker, Chris Mason (the author of Mercurial Queues), wrote a similar tool called mpatch [Mas], which takes a simple approach to automating the application of hunks rejected by patch. The mpatch command can help with four common reasons that a hunk may be rejected:

• The context in the middle of a hunk has changed.
• A hunk is missing some context at the beginning or end.
• A large hunk might apply better—either entirely or in part—if it was broken up into smaller hunks.
• A hunk removes lines with slightly different content than those currently present in the file.

If you use wiggle or mpatch, you should be doubly careful to check your results when you’re done. In fact, mpatch enforces this method of double-checking the tool’s output, by automatically dropping you into a merge program when it has done its job, so that you can verify its work and finish off any remaining merges.

12.7 Getting the best performance out of MQ

MQ is very efficient at handling a large number of patches. I ran some performance experiments in mid-2006 for a talk that I gave at the 2006 EuroPython conference [O’S06]. I used as my data set the Linux 2.6.17-mm1 patch series, which consists of 1,738 patches. I applied these on top of a Linux kernel repository containing all 27,472 revisions between Linux 2.6.12-rc2 and Linux 2.6.17.

On my old, slow laptop, I was able to “hg qpush -a” all 1,738 patches in 3.5 minutes, and “hg qpop -a” them all in 30 seconds. (On a newer laptop, the time to push all patches dropped to two minutes.) I could “hg qrefresh” one of the biggest patches (which made 22,779 lines of changes to 287 files) in 6.6 seconds.

Clearly, MQ is well suited to working in large trees, but there are a few tricks you can use to get the best performance of it.

First of all, try to “batch” operations together. Every time you run “hg qpush” or “hg qpop”, these commands scan the working directory once to make sure you haven’t made some changes and then forgotten to run “hg qrefresh”. On a small tree, the time that this scan takes is unnoticeable. However, on a medium-sized tree (containing tens of thousands of files), it can take a second or more.

The “hg qpush” and “hg qpop” commands allow you to push and pop multiple patches at a time. You can identify the “destination patch” that you want to end up at. When you “hg qpush” with a destination specified, it will push patches until that patch is at the top of the applied stack. When you “hg qpop” to a destination, MQ will pop patches until the destination patch is at the top.

You can identify a destination patch using either the name of the patch, or by number. If you use numeric addressing, patches are counted from zero; this means that the first patch is zero, the second is one, and so on.

12.8 Updating your patches when the underlying code changes

It’s common to have a stack of patches on top of an underlying repository that you don’t modify directly. If you’re working on changes to third-party code, or on a feature that is taking longer to develop than the rate of change of the code beneath, you will often need to sync up with the underlying code, and fix up any hunks in your patches that no longer apply. This is called rebasing your patch series.

The simplest way to do this is to “hg qpop -a” your patches, then “hg pull” changes into the underlying repository, and finally “hg qpush -a” your patches again. MQ will stop pushing any time it runs across a patch that fails to apply during conflicts, allowing you to fix your conflicts, “hg qrefresh” the affected patch, and continue pushing until you have fixed your entire stack.

This approach is easy to use and works well if you don’t expect changes to the underlying code to affect how well your patches apply. If your patch stack touches code that is modified frequently or invasively in the underlying repository, however, fixing up rejected hunks by hand quickly becomes tiresome.

It’s possible to partially automate the rebasing process. If your patches apply cleanly against some revision of the underlying repo, MQ can use this information to help you to resolve conflicts between your patches and a different revision.

The process is a little involved.

1. To begin, “hg qpush -a” all of your patches on top of the revision where you know that they apply cleanly.
2. Save a backup copy of your patch directory using “hg qsave -e -c”. This prints the name of the directory that it has saved the patches in. It will save the patches to a directory called .hg/patches.N, where N is a small integer. It also commits a “save changeset” on top of your applied patches; this is for internal book-keeping, and records the states of the series and status files.
3. Use “hg pull” to bring new changes into the underlying repository. (Don’t run “hg pull -u”; see below for why.)
4. Update to the new tip revision, using “hg update -C” to override the patches you have pushed.
5. Merge all patches using “hg qpush -m -a”. The -m option to “hg qpush” tells MQ to perform a three-way merge if the patch fails to apply.

During the “hg qpush -m”, each patch in the series file is applied normally. If a patch applies with fuzz or rejects, MQ looks at the queue you “hg qsave”d, and performs a three-way merge with the corresponding changeset. This merge uses Mercurial’s normal merge machinery, so it may pop up a GUI merge tool to help you to resolve problems.

When you finish resolving the effects of a patch, MQ refreshes your patch based on the result of the merge.

At the end of this process, your repository will have one extra head from the old patch queue, and a copy of the old patch queue will be in .hg/patches.N. You can remove the extra head using “hg qpop -a -n patches.N” or “hg strip”. You can delete .hg/patches.N once you are sure that you no longer need it as a backup.

12.9 Identifying patches

MQ commands that work with patches let you refer to a patch either by using its name or by a number. By name is obvious enough; pass the name foo.patch to “hg qpush”, for example, and it will push patches until foo.patch is applied.

As a shortcut, you can refer to a patch using both a name and a numeric offset; foo.patch-2 means “two patches before foo.patch”, while bar.patch+4 means “four patches after bar.patch”.

Referring to a patch by index isn’t much different. The first patch printed in the output of “hg qseries” is patch zero (yes, it’s one of those start-at-zero counting systems); the second is patch one; and so on

MQ also makes it easy to work with patches when you are using normal Mercurial commands. Every command that accepts a changeset ID will also accept the name of an applied patch. MQ augments the tags normally in the repository with an eponymous one for each applied patch. In addition, the special tags qbase and qtip identify the “bottom-most” and topmost applied patches, respectively.

These additions to Mercurial’s normal tagging capabilities make dealing with patches even more of a breeze.

• Want to patchbomb a mailing list with your latest series of changes?
  1  hg email qbase:qtip

  (Don’t know what “patchbombing” is? See section 14.4.)

• Need to see all of the patches since foo.patch that have touched files in a subdirectory of your tree?
  1  hg log -r foo.patch:qtip subdir

Because MQ makes the names of patches available to the rest of Mercurial through its normal internal tag machinery, you don’t need to type in the entire name of a patch when you want to identify it by name.