[DRAFT] Branching and merging with git
From: <hidden>
Date: 2016-08-11 19:54:31
I know it took me a while to get used to playing with branches, and I
still get nervous when doing something creative. So I've been trying
to get more comfortable, and wrote the following to document what I've
learned.
It's a first draft - I just finished writing it, so there are probably
some glaring errors - but I thought it might be of interest anyway.
* Branching and merging in git
In CVS, branches are difficult and awkward to use, and generally
considered an advanced technique. Many people use CVS for a long time
without departing from the trunk.
Git is very different. Branching and merging are central to effective use
of git, and if you aren't comfortable with them, you won't be comfortable
with git. In particular, they are required to share work with other
people.
The only things that are a bit confusing are some of the names.
In particular, at least when beginning:
- You create new branches with "git checkout -b".
"git branch" should only be used to list and delete branches.
- You share work with "git fetch" and "git push". These are opposites.
- You merge with "git pull", not "git merge". "git pull" can
also do a "git fetch", but that's optional. What's not optional
is the merge.
* A brief digression on command names.
Originally, all git commands were named "git-foo". When there got to
be over a hundred, people started complaining about the clutter in
/usr/bin. After some discussion, the following solution was reached:
- It's now possible to place all of the git-foo commands into a separate
directory. (Despite the complaints, not too many people are doing it
yet.)
- One option for git users is to add that directory to their $PATH.
- Another is provided by a wrapper called just "git". It's intended to
live in a public directory like /usr/bin, and knows the location of
the separate directory. When you type "git foo", it finds and executes
"git-foo".
- Some simple commands are built into the git wrapper. When you type
"git add", it just does it internally. (On the git mailing list,
you will see patches like "make git diff a builtin"; this is what
they're talking about.)
- For compatibility, for each builtin, there is a "git-add" file,
which is just a link to the "git" wrapper. It looks at the name it
was invoked as to figure out what it should do.
The one confusing thing is that, although people usually type "git foo"
in examples, they're interchangeable in practice. I go back and forth
for no good reason. The main caveat is that to get the man page, you
still need to type "man git-foo". Fortunately, there are two other ways
to get the man page:
1) "git help foo"
2) "git foo --help"
Git doesn't have a specialized built-in help system; it just shows you
the man pages.
One outstanding problem with git's man pages is that often the most detail
is in the command page that was written first, not the user-friendly
one that you should use. For example, there are a number of special
cases of the "git diff" command that were written first, and the man
pages for these commands (git-diff-index, git-diff-files, git-diff-tree,
and git-diff-stages) are considerably more informative than the page for
plain git-diff, even though that's the command that you should use 99%
of the time.
* Git's representation of history
As you recall from Git 101, there are exactly four kinds of objects in
Git's object database. All of them have globally unique 40-character hex
names made by hashing their type and contents. Blob objects record file
contents; they contain bytes. Tree objects record directory contents;
they contain file names, permissions, and the associated tree or blob
object names. Tag objects are shareable pointers to other objects;
they're generally used to store a digital signature.
And then, we come to commit objects. Every commit points to (contains
the name of) an associated tree object which records the state of the
source code at the time of the commit, and some descriptive data (time,
author, committer, commit comment) about the commit.
And most importantly, it contains a list of "parent commits", older
commits from which this one is derived. These pointers are what produce
the history graph.
Typically only one commit (the initial commit) has zero parents. It's
possible to have more than one such commit (if you merge two projects
with different history), but that's unusual.
Many commits have exactly one parent. These are made by a normal commit
after editing. From a branching and merging point of view, they're not
too exciting.
And then there are commits which have multiple parents. Two is most
common, but git allows many more. (There's a limit of sixteen in the
source code, and the most anyone's ever used in real life is 12, and
that was generally regarded as overdoing it. Google on "doedecapus"
for discussion of it.)
Finally, there are references, stored in the .git/refs directory.
These are the human-readable names associated with commits, and the
"root set" from which all other commits should be reachable.
These references are generally divided into two types, although
there is no fundamental difference:
- Tags are references that are intended to be immutable.
The "v1.2" tag is a historical record. A tag may point to
a tag object (which will hold a signature), or just to a commit
directly. The latter isn't cryptographically authenticated, but
works just fine for everyday use.
- Heads are references that are intended to be updated. "Head"
is actually synonymous with "branch", although one emphasizes the
tip more, while the other directs your attention to the entire
path that got there.
Either way, they're just a 41-byte file that contains a 40-byte hex
object ID, plus a newline. Tags are stored in .git/refs/tags, and heads
are stored in .git/refs/heads. Creating a new branch is literally just
picking a file name and writing the ID of an existing commit into it.
The git programs enforce the immutability of tags, but that's a safety
feature, not something fundamental. You can rename a tag to the heads
directory and go wild.
The only limit on branches is clutter. A number of git commands have
ways to operate on "all heads", and if you have too many, it can get
annoying. If you're not using a branch, either delete it, or move it
somewhere (like the tags directory) where it won't clutter up the list of
"currently active heads".
(Note that CVS doesn't have this all-heads default, so people tend to
use longer branch names and keep them around after they've been merged
into the trunk. Old CVS repositories converted to git generally need
an old-branch cleanup.)
Another thing that's worth mentioning is that head and tag names can
contain slashes; i.e. you're allowed to make subdirectories in the
.git/refs/heads and .git/refs/tags directories. See the name page
for "git-check-ref-format" for full details of legal names.
* Naming revisions
CVS encourages you to tag like crazy, because the only other way to
find a given revision is by date. Git makes it a lot easier, so most
revisions don't need names.
You can find a full description in the git-rev-parse man page, but here's
a summary.
First of all, every commit has a globally unique name, its 40-digit hex
object ID. It's a bit long and awkward, but always works. This is useful
for talking about a specific commit on a mailing list. You can abbreviate
it to a unique prefix; most people find about 8 digits sufficient.
(Subversion is easier yet, because it assigns a sequential number to each
commit. However, that isn't possible in a distributed system like git.)
Second, you can refer to a head or tag name. Git looks in the
following places, in order, for a head:
1) .git
2) .git/refs
3) .git/refs/heads
4) .git/refs/tags
You should avoid having e.g. a head and a tag with the same name, but
if you do, you can specify one or the other with heads/foo and tags/foo.
Third, you can specify a commit relative to another. The simplest
one is "the parent", specified by appending ^ to a name. E.g. HEAD^
or deadbeef^. If there are multiple parents, then ^ is the same as ^1,
and the others are ^2, ^3, etc.
So the last few commits you've made are HEAD, HEAD^, HEAD^^, HEAD^^^, etc.
After a while, counting carets becomes annoying, so you can abbreviate
^^^^ as ~4. Note that this only lets you specify the first parent.
If you want to follow a side branch, you have to specify something like
"master~305^2~22".
* Converting between names
Git has two helpers (programs designed mainly for use in shell scripts)
to convert between global object IDs and human-readable names.
The first is git-rev-parse. This is a general git shell script helper,
which validates the command line and converts object names to absolute
object IDs. Its man page has a detailed description of the object
name syntax.
The second is git-name-rev, which converts the other way around. It's
particularly useful for seeing which tags a given commit falls between.
* Working with branches, the trivial cases.
By convention, the local "trunk" of git development is called "master".
This is just the name of the branch it creates when you start an empty
repository. You can delete it if you don't like the name.
If you create your repository by cloning someone else's repository, the
remote "master" branch is copied to a local branch named "origin". You
get your own "master" branch which is not tied to the remote repository.
There is always a current head, known as HEAD. (This is actually a
symbolic link, .git/HEAD, to a file like refs/heads/master.) Git requires
that this always point to the refs/heads directory.
Minor technical details:
1) HEAD used to be a Unix symlink, and can still be though of that
way, but for Microsoft support, this is now what's called a
"symbolic reference" or symref, and is a plain file containing
"ref: refs/heads/master". Git treats it just like a symlink.
There's a git-update-ref helper which writes these.
2) While HEAD must point to refs/heads, it's legal for it to
point to a file that doesn't exist. This is what happens
before the first commit in a brand new repository.
When you do "git commit", a new commit object is created with the old
HEAD as a parent, and the new commit is written to the current head
(pointed to by HEAD).
* The three uses of "git checkout"
Git checkout can do three separate things:
1) Change to a new head
git checkout [-f|-m] <branch>
This makes <branch> the new HEAD, and copies its state to the index
and the working directory.
If a file has unsaved changes in the working directory, this tries
to preserve them. This is a simple attempt, and requires that the
modified files(s) are not altered between the old and new HEADs.
In that case, the version in the working directory is left untouched.
A more aggressive option is -m, which will try to do a three-way
(intra-file) merge. This can fail, leaving unmerged files in the
index.
An alternative is to use -f, which will overwrite any unsaved changes
in the working directory. This option can be used with no <branch>
specified (defaults to HEAD) to undo local edits.
2) Revert changes to a small number of files.
git checkout [<revision>] [--] <paths>
will copy the version of the <paths> from the index to the working
directory. If a <revision> is given, the index for those paths will
be updated from the given revision before copying from the index to
the working tree.
Unlike the version with no <paths> specified, this does NOT update
HEAD, even if <paths> is ".".
3) Create a branch.
git checkout [-f|-m] -b <branch> [revision]
will create, and switch to, a new branch with the given name.
This is equivalent to
git branch <branch> [<revision>]
git checkout [-f|-m] <branch>
If <revision> is omitted, it defaults to the current HEAD, in which
case no working directory files are altered.
This is the usual way that one checks out a revision that does not
have an existing head pointing to it.
* Deleting branches
"git branch -d <head>" is safe. It deletes the given <head>, but first
it checks that the commit is reachable some other way. That is, you
merged the branch in somewhere, or you never did any edits on that branch.
It's a good idea to create a "topic branch" when you're working on
anything bigger than a one-liner, but it's also a good idea to delete
them when you're done. It's still there in the history.
* Doing rude things to heads: git reset
If you need to overwrite the current HEAD for some reason, the tool to
do it with is "git reset". There are three levels of reset:
git reset --soft <head>
This overwrites the current HEAD with the contents of <head>.
If you omit <head>, it defaults to HEAD, so this does nothing.
git reset [<head>]
git reset --mixed [<head>]
These overwrite the current HEAD, and copy it to the index,
undoing any git-update-index commands you may have executed.
If you omit <head>, it default to HEAD, so there is no change
to the current branch, but all index changes are undone.
git reset --hard [<head>]
This does everything mentioned above, and updates the
working directory. This throws away all of your in-progress
edits and gets you a clean copy. This is also commonly
used without an explicit <head>, in which case the current
HEAD is used.
* Using git-reset to fix mistakes
"Oh, no! I didn't mean to commit *that*! How do I undo it?"
If you just want to undo a commit, then you can use "git reset HEAD^"
to return the current HEAD to the previous version. If you want to leave
the commit in the index (this only applies to you if you are familiar with
using the index; see below), then you can use "git reset --soft HEAD^".
And if you want to blow away every record of the changes you made,
you can use "git reset --hard HEAD^"
If you just want a stupid trivial mistake and want to replace the most
recent commit with a corrected one, "git commit --amend" is your friend.
It makes a new commit with HEAD^ rather than HEAD as its ancestor.
* Fixing mistakes without git-reset
git-reset has the problem that it doesn't preserve hacking in progress
in the working directory. It can leave the working directory alone
(making everything a "hack in progress"), but it can't merge changes
like git checkout.
So, suppose you've been trying something that should have been simple, and
made three commits before realizing that the problem is harder than you
thought and you want your work so far to be on a new branch of its own;
committing them on the current HEAD (I'll call it "old") was a mistake.
You don't want to erase anything, just rename it. Make "new" a copy of
the current "old" and move old back to HEAD^^^ (three commits ago).
While there are ways to do that using git-reset, but far better is
to use "git branch -f":
git checkout -b new
Create (and switch to) the "new" branch.
git branch -f old HEAD^^^
Forcibly move "old" back three versions.
(You could also use old~3 or new^^^ or any synonymous name.)
You can use a similar trick to rename a branch. If it's the current
HEAD, then:
git checkout -b newname
git branch -d oldname
and if it's not, then
git branch newname oldname
git branch -d oldname
An alternative in the latter case is to just use mv on the raw
.git/refs/heads/oldname file.
* How do I check out an old version?
A very common beginning question is how to check out an old version.
Say you need to compile an old release for test purposes. "git checkout
v1.2" gives a funny error message. What's going on?
Well, "git checkout" makes the current HEAD point to the head that
you specify. And, as previously mentioned, git requires that it point
to something in the .git/refs/heads directory. So you can't do that.
If you're busy doing things in your working directory, and don't want to
overwrite your work with an old version, then you can get a snapshot with
the (old) git-tar-tree or (new) git-archive commands. These produce a
tar file (git-archive can also produce a zip file) which is a snapshot
of any version you like. You can then unpack this file in a different
directory and build it.
However, if you haven't got any edits in progress, and want to check out
the old version into your working directory, just create a temp branch!
git checkout -b temp v1.2
Will do what you want. This will also do what you want if you have a
local edit (like the "#define DEBUG 1" mentioned above) that you want
to preserve while working on the old version.
You'll see this in use if you ever use the (highly recommended) git-bisect
tool. It creates a branch called "bisect" for the duration of the bisect.
(Yes, I have to confess, I sometimes wish that git would enforce the
"HEAD must point to .git/refs/heads" rule when committing (checking in)
rather than when checking out, but that's the way git has grown up.)
Note that if you want *exactly* an old version, with no local hacks,
make sure there are none (with "git status") when doing this. It's more
convenient if you do it before the checkout, but you'll get the same
answer if you ask afterwards.
Now, what about the complex case: you have local hacks that you
want to keep, but not have polluting the old version?
Well, one way of the other, you'll have to commit it. If you don't mind
committing your changes to the current branch ("git commit -a"), do that.
If they're not ready to commit, you can commit them anyway, and back
them out when you're done:
git commit -a -m "Temp commit"
git checkout -b temp v1.2
make ; make test ; whatever
git checkout master
git branch -d temp
git reset HEAD^
This leaves both the working directory and the master head in the states
they were in at the beginning.
If you don't like committing to the master branch, you can make a new one.
In this example, it's "work in progress", a.k.a. "wip":
git checkout -b wip
git commit -a -m "Temp commit"
git checkout -b temp v1.2
make ; make test ; whatever
git checkout wip
git branch -d temp
git reset master
git checkout master # Won't change working directory
git branch -d wip
* Examining history: git-log and git-rev-list
In another example of docs being better on the first command written,
the all-purpose utility for examining history is "git log", but all of
the examples of clever ways to use it are in the git-rev-list man page.
And git-log also has most of git-diff's options.
Other utilities, notably the gitk and qgit GUIs, also use the git-rev-list
command-line options, so it's well worth learning them.
git-rev-list gives you a filtered subset of the repository history.
There are two basic ways that you can do the filtering:
1) By ancestry. You specify a set of commits to include all the
ancestors of, and another set to exclude all the ancestors of.
(For this purpose, a commit is considered an ancestor of itself.)
So if you want to see all commits between v1.1 and v1.2, you
can specify
git log ^v1.1 v1.2
or, with a more convenient syntax
git log v1.1..v1.2
However, there are times when you want to specify something more
complex. For example, if a big branch that had been in progress since
v1.0.7 was merged between v1.1 and v1.2, but you don't want to see it,
you could specify any of:
git log v1.2 ^v1.1 ^bigbranch
git log ^bigbranch v1.1..v1.2
git log ^v1.1 bigbranch..v1.2
They're all equivalent. Another special syntax that's sometimes
handy is
git log branch1...branch2
Note the three dots. This generates the symmetric difference between
the two; basically it's a diff between the commits that went into
each of them.
"git log" by default pipes its output through less(1), and generates
its output from newest to oldest on the fly, so there's no great
speed penalty to not specifying a starting place. It'll generate a
few screen fulls more than you look at, but not waste any more effort
than that.
2) By path name. This is a feature which appears to be unique to git.
If you give git-rev-list (or git-log, or gitk, or qgit) a list of
pathname prefixes, it will list only commits which touch those
paths. So "git log drivers/scsi include/scsi" will list only
commits which alters a file whose name begins with drivers/scsi
or include/scsi.
(If there's any possible ambiguity between a path name and a commit
name, git-rev-list will refuse to proceed. You can resolve it by
including "--" on the command line. Everything before that is a
commit name; everything after is a path.)
This filter is in addition to the ancestry filter. It's also rather
clever about omitting unnecessary detail. In particular, if there's
a side branch which does touch drivers/scsi, then the entire branch,
and the merge at the end, will be removed from the log.
You can additionally limit the commits to a certain number, or by date,
author, committer, and so on.
By default, "git log" only shows the commit messages, so it's important to
write good ones. Other tools compress commit messages down to
the first line, so try to make that as informative as possible.
* History diagrams
When talking about various situations involving multiple branches,
people often find it handy to draw pictures. Gitk draws nice pictures
vertically, but for e-mail, ASCII art drawn horizontally is often easier.
Commits are shown as "o", and the links between them with lines drawn with
- / and \. Time goes left to right, and heads may be labelled with names.
For example:
o--o--o <-- Branch A
/
o--o--o <-- master
\
o--o--o <-- Branch B
If someone needs to talk about a particular commit, the character "o"
may be replaced with another letter or number.
* Trivial merges: fast-forward and already up-to-date.
There are two kinds of merge that are particularly simple, and you will
encounter them in git a great deal. They are mirror images.
Suppose that you are working on branch A and merge in branch B, but no
work has been done to branch B since the last time you merged, or since
you spawned branch A from it. That is, the history looks like
o--o--o--o <-- B
\
o--o--o <-- A
or
o--o--o--o--o--o <-- B
\ \
o--o--o--o--o <-- A
If you then merge B into A, A is described as "already up to date".
It is already a strict superset of B, and the merge does nothing.
In particular, git will not create a dummy commit to record the fact that
a merge was done. It turns out that are a number of bad things that would
happen if you did this, but for now, I'll just say that git doesn't do it.
Now, the opposite scenario is the "fast-forward" merge. Suppose you
merge A into B. Again, A is a strict superset of B.
In this case, git will simply change the head B to point to the same
commit as A and say that it did a "fast-forward" merge. Again, no commit
object is created to reflect this fact.
The effect is to unclutter the git history. If I create a topic branch to
work on a feature, do some hacking, and then merge the result back into
the (untouched!) master, the history will look just like I did all the
work on the master directly. If I then delete the topic branch (because
I'm done using it), the repository state is truly indistinguishable.
While the topic branch existed, you could have done something to the
master branch, in which case the final merge would have been non-trivial,
but if that didn't happen, git produces a simple, easy-to-follow linear
history.
Some people used to heavyweight branches find this confusing; they
think a merge is a big deal and it should be memorialized, but there
are actually excellent reasons for doing this.
The most important one is that a fit of merging back and forth will
eventually end. Suppose that branches A and B are maintained by separate
developers who like to track each other's work closely.
If the fast-forward case did create a commit, then merging A into B
would produce
o--o--o--o---------o <-- B
\ /
o--o--o <-- A
then merging B into A would produce:
o--o--o--o---------o <-- B
\ / \
o--o--o---o <-- A
and further merges would produce more and more dummy commits, all without
ever reaching a steady state, and without making it obvious that the
two heads are actually identical.
Since history lasts forever, cluttering it up with unimportant stuff is a
burden to all future users, and not a good idea. Allowing the merge of a
branch to be seamless in the simple case encourages lightweight branches.
If you _might_ need a separate branch, create it. If it turned out that
you didn't, it won't make a difference.
* Exchanging work with other repositories
The basic tools for exchanging work with other repositories are "git
fetch" and "git push". The fact that "git pull" is not the opposite of
"git push" is often confusing to beginners (it's a superset of git fetch),
but that's the terminology that has grown up.
The unit of sharing in git is the branch. If you've used branches in
CVS, you'll be familiar with using "CVS update" to pull changes from your
"current branch" in the repository into your working directory.
In Git, you don't pull into the working directory, but rather into a
tracking branch. You set up a branch in your repository which will be
a copy of the branch in the remote repository. For example, if you use
"git clone", then the remote "master" branch is tracked by the local
"origin" branch.
Then, when you do a "git fetch", git fetches all of the new commits
and sets the origin head to point to the newly fetched head of the
remote branch.
By default, git checks that this is a trivial fast-forward merge, that
is not throwing away history. If it finds something like:
o--o--o--o--o--o <-- remote master
\
o <-- Local origin
It will complain and abort the fetch. This is usually a warning that
something has gone wrong - in particular, you forgot that this was
supposed to be a tracking branch and committed some work to it - and it
aborts before throwing your work away.
However, sometimes the remote git user will have a branch name that they
delete and re-create frequently. There are plenty of reasons to do this.
The most common is doing a "test merge" between various branches in
progress. They're all unfinished, so the developer of branch A doesn't
want to merge in all the new bugs in branch B, but a tester might want
to create a merged version with both sets of bugs for testing.
The merged version is not intended to be a permanent part of history -
it'll get deleted after the test - but it can still be useful to have
a draft copy.
In this case, you can mark the source branch with a leading "+", to
disable this sanity check. (See the git-fetch man page for details.)
Note that in this case, you should specifically avoid merging from such
a branch into any non-test branches of your own. It is, as mentioned,
not intended to be a permanent part of history, so don't make it part
of your permanent history. (You still might want to test-merge it with
your work in progress, of course.)
The fact that you should know to treat such branches specially is why
git doesn't try to automatically cope with them.
* Alternate branch naming
The original git scheme mixes tracking branches with all the other heads.
This requires that you remember which branches are tracking branches and
which aren't. Hopefully, you remember what all your branches are for,
but if you track a lot of remote repositories, you might not remember
what every remote branch is for and what you called it locally.
* Remotes files
You can specify what to fetch on the git-fetch command line. However,
if you intend to monitor another repository on an ongoing basis,
it's generally easier to set up a short-cut by placing the options in
.git/remotes/<name>.
The syntax is explained in the git-fetch man page. When this is st
up, "git fetch <name>" will retrieve all the branches listed in the
.git/remotes/<name> file. The ability to fetch multiple branches at
once (such as release, beta, and development) is an advantage of using
a remotes file.
You can also create the remotes file "origin" (not necessarily any
relation to the branch named "origin"), which is the default for
git-fetch. If you have a single primary "upstream" repository that
you sync to, place it in the origin remotes file, and you can just type
"git fetch" to get all the latest changes.
Note that branches to fetch are identified by "Pull: " lines in the
remotes file. This is another example of the fetch/pull confusion.
git-pull will be explained eventually.
* Remote tags
TODO: Figure out how remote tags work, under what circumstances
they are fetched, and what git does if there are conflicts.
* Exchanging work with other repositories, part II: git-push
It's simpler to set up git sharing on a pull basis. If your source
code isn't secret, you can set up a public read-only server very easily
(see the git-daemon man page for details), and have other fetch from that.
However, N developers all pulling from each other is an N^2 mess.
Some centralization helps.
One way is to have a central coordinator (like Linus) who pulls from
all of the developers, and who they in turn pull from.
The other is to have a central repository that people can push to.
This generally requires an ssh login on the server. You can use git-shell
as the login shell if all you want to allow the account to do is git
fetch and push. (You can use the hook scripts to enforce rules about
who's allowed to do what to which branch.)
Git-push to the remote machine works exactly like git-fetch from the
remote machine. The objects are moved over, and the branches pushed to
are fast-forwarded. If fast-forward is impossible, you get an error.
So if you have multiple people committing to a branch on the server,
you will not be allowed to push if someone has pushed more to that branch
since last time you fetched it.
You have to merge the changes locally, and re-try the push when you've
got a new head that includes the most recently pushed work as an ancestor.
This is exactly like "cvs commit" not working if your recent checkout
wasn't the (current) tip of the branch, but git can upload more than
one commit.
The simplest way to resolve the conflict is to merge the remote head with
your local head. This is easiest if you have different local branches
for fetching the remote repository and for pushing to it.
That is, you have one head that just tracks the master repository's
main branch, and another that you add your work to, and push from.
This makes merging simpler when there are conflicts.
Another use for git-push, even for a solo developer, is sharing your work
with the world. You can set up a public git server on a high-bandwidth
machine (possibly rented from a hosting service) and then push to it to
publish something.
* Merging (finally!)
I went through everything else first because the most common merge case
is local changes with remote changes. Not that you can't merge two
branches of your own, but you don't need to do that nearly as often.
The primitive that does the merging is called (guess what?) git-merge.
And you can use that if you want. If you want to create a so-called
octopus merge, with more than two parents, you have to.
However, it's usually easier to use the git-pull wrapper. This merges
the changes from some other branch into the current HEAD and generates
a commit message automatically.
git-merge lets you specify the commit message (rather than generating it
automatically) and use a non-HEAD destination branch, but those options
are usually more annoying than useful.
The basic git-pull syntax is
git-pull <repository> <branch>
The repository can be any URL that git supports. Including, particularly,
a local file. So to do a simple local merge, you just type
git-pull . <branch>
So after doing some hacking on branch "foo", you would
git checkout master
git pull . foo
and ba-boom, all is done.
Now, you can also specify a remote repository to merge from, using a
git://, http:// or git+ssh:// URL. This is what Linus does all day long,
and why the git-pull tool is optimized to allow that. It uses git-fetch
to fetch the remote branch without assigning it a branch name (it gets
the special name FETCH_HEAD temporarily), and them merges it into the
current HEAD directly.
There is absolutely nothing wrong with doing that, but beginners often
find it confusing to have a single short command do quite so much.
And if you are working closely with someone, it's often more convenient
and less confusing to keep local tracking branches. Then you can
git fetch upstream # Fetches 'origin'
git pull . origin
It's also possible to give just a single remotes file name to git-pull:
git pull upstream
That does a git fetch, updating all of the listed branches as usual,
then merges the _first_ listed branch into HEAD.
By the way: don't blink, you might miss it! As I mentioned, pulling is
a very big part of Linus's daily routine, and he's made sure it's fast.
(Actually, it produces a fair bit of output, so you'll see.)
Just to clarify, because people often get confused:
git-pull is a MERGING tool. It always does a merge, as well as an optional
fetch. If you just want to LOOK at a remote branch, use git-fetch.
* Undoing a merge
If you discover that a merge was a mistake, it can be undone just like
any other commit. The HEAD you merged to is the first parent, so just do
git reset --hard HEAD^
This is why Linus likes a git-pull command that does so much in one shot -
if he doesn't like what he pulls, it's easy to undo.
* How merging operates
Git uses the basic three-way merge. First, it applies it to whole files,
and then to lines within files.
To do a three-way merge, you need three versions of a file. The versions
A and B you want to merge, and a common ancestor, commonly called O.
That is, history proceeds something like:
o--o--A
/
o--o--O
\
o--B
The basic idea is "I want the file O, plus all the changes made from O
to A, plus all the changes made from O to B." Since the cases where one
of A or B is a direct ancestor of the other have already been disposed
of, the three commits must be different.
For each file, there are a few cases that are trivial, and git gets
these out of the way immediately:
- If A and B are identical, the merged result is obvious.
- If O and A are the same, then the result should be B.
- If O and B are the same, then the result should be A.
In the completely trivial case when O, A and B are the same, then
all three rules apply, they all produce the same obvious result.
The "merge base" version O is generally the most recent common ancestor
of A and B. The only problem is, that's not necessarily unique!
The classic confusing case is called a "criss-cross merge", and looks
like this:
o--b-o-o--B
/ \ /
o--o--o X
\ / \
o--a-o-o--A
There are two common ancestors of A and B, marked a and b in the graph
above. And they're not the same. You could use either one and get
reasonable results, but how to choose?
The details are too advanced for this discussion, but the default
"recursive" merge strategy that git uses solves the answer by merging
a and b into a temporary commit and using *that* as the merge base.
Of course, a and b could have the same problem, so merging them could
require another merge of still-older commits. This is why the algorithm
is called "recursive." It's been tested with pathological conditions,
but multiply nested criss-cross merges are very rare, so the recursion
isn't a performance limit in practice.
If all three of a given file in O, A, B are different, then the three
versions are pulled into the index file, called "stage 1", "stage 2",
and "stage 3", and a merge strategy driver is called to resolve the mess.
Git then uses the classic line-based three-way merge, looking for isolated
changes and applying the same rules as for files when two of the source
files are the same in some range.
* Alternate merge strategies
In every version control system prior to git, the merging algorithm was
buried deep in the bowels of the software, and very difficult to change.
One of particularly nice things that git did was allow for easily
replaceable "merge strategies". Indeed, you can try multiple merge
strategies, and the fallback - print an error message and let the user
sort it out - can be thought of as just another merge strategy.
Enabling this is why the index is so important to git. It provides a
place to store an unfinished merge, so you can try various strategies
(including hand-editing) to finish it.
Generally, git's default merge strategies are just fine. There is,
however, one special case that is occasionally useful, specified with the
"-s ours" strategy.
That strategy instructs git that the merged result should be the same
as the current HEAD. Any other branches are recorded as parents, but
their contents are ignored.
What the heck is the use of that? Well, it lets you record the fact
that some work has been done in the history, and that it shouldn't be
merged again. For example, say you write and share a popular patch set.
People are always merging it in to their local source trees. But then
you discover a much better way to achieve the goal of that patch set, and
you want to publish the fact that the new patch supersedes the old one.
If you developed the new set starting from the old one, that would happen
automatically. But another way to achieve the same goal is to merge the
old branch it in using the "ours" strategy. Everyone else's git will
notice that the patch is already included, and stop trying to merge it in.
* When merging goes wrong
This is the fun part. Git's default recursive-merge strategy is pretty
clever, but sometimes changes truly do conflict and need manual fix-up.
When git is unable to complete a merge, it leaves the three different
versions in the index and places a file with CVS-style conflict markers
in the working directory.
As long as there is a "staged" file in the index, you will not be able
to commit. You must resolve the conflict, and update the index with the
resolved versions. You can do this one at a time with git-update-index,
or at the end by giving the files as arguments to git-commit.
Doing them one at a time is probably safest; checking in a file which still
has conflict markers makes a bit of a mess. Note that git will still
use the automatically generated commit message when you finally commit.
(It's in .git/MERGE_MSG, if you care.)
Note that "git diff" knows how to be useful with a staged file.
By default, it displays a multi-way diff. For example, suppose I take a
(slightly buggy) hello.c:
--- hello.c ---#include <stdio.h>
int main(void)
{
printf("Hello, world!");
}--- end ---
Now, suppose that in branch A, I fix some bugs - add the missing newline and "return 0;". In branch B, I display my angst and change it to "Goodbye, cruel world!". When I try to merge A into B, obviously I'll get a conflict. The resultant file, with conflict markers, looks like:
--- hello.c ---#include <stdio.h>
int
main(void)
{
<<<<<<< HEAD/hello.c
printf("Goodbye, cruel world!");
=======
printf("Hello, world!\n");
return 0;quoted
quoted
quoted
quoted
quoted
quoted
edadc53fc7a8aef2a672a4fa9d09aa16f4e14706/hello.c
}
--- end ---
and the result of "git diff" is diff --cc hello.c index 4b7f550,948a5f8..0000000
--- a/hello.c
+++ b/hello.c@@@ -3,5 -3,6 +3,10 @@@
int
main(void)
{
++<<<<<<< HEAD/hello.c
+ printf("Goodbye, cruel world!");
++=======
+ printf("Hello, world!\n");
+ return 0;
++>>>>>>> edadc53fc7a8aef2a672a4fa9d09aa16f4e14706/hello.c
}
Notice how this is not a standard diff! It has two columns of diff
symbols, and shows the difference from each of the ancestors to the
current hello.c contents. I can also use "git diff -1" to compare
against the common ancestor, or "-2" or "-3" to compare against each of
the merged copies individually.
* Alternatives to merging
The bigger and more active your source tree, the more important it is to
keep the history reasonably clean. Just because git can do a merge in
under a second doesn't mean that you should do one daily. When you look
back at a feature's development history, you'd like to see meaningful
changes recorded and not a lot of meaningless ones.
Now, once you have shared a commit with others, and they have incorporated
it into their development, it becomes impossible to undo. But git
provides tools that are useful for "rewriting history" before public
release. These can be used to edit a commit for publication.
* Test merging
One way to keep the history clean is to simply not merge other branches
into your development branch. If you want to use your new features and
other people's code changes, make a test merge and use that, but don't
make that merge part of your branch.
This is slightly more work (you have to change to a test branch and do
your merging there), but not very much.
Sometimes, when doing this, a conflict appears between your changes and
someone else's development. If you get tired of fixing the same conflict
every time you do a test merge, have a look at the git-rerere tool.
This remembers resolved conflicts and tries to apply the same resolution
patch the next time.
It's written specifically to help you not do an extra merge unnecessarily.
Although its man page is well worth reading, you never invoke git-rerere
explicitly; it's invoked automatically by the merge and patch tools if
you create a .git/rr-cache directory.
* Cherry picking
If you have a series of patches on a branch, but you want a subset
of them, or in a different order, there's a handy utility called
"git-cherry-pick" which will find the diff and apply it as a patch to
the current HEAD. It automatically recycles the commit message from
the original commit.
If the patch can't be applied, it leaves the versions in the index and
conflict markers in the working directory just like a failed merge.
And just like a merge, it remembers the commit message and provides it
as a default when I finally commit.
Note that this can only work on a chain of single-parent commits.
If a commit has multiple parents, there's no single patch to apply.
You van get the list of commits on a branch with git-log or git-rev-list,
but for more complex cases, the git-cherry tool is designed to generate
the list of commits to merge. It has a rather neat approximate-match
function built in which identifies patches that appear to already be
present in the target branch.
* Rebasing
A special case of cherry-picking is if you want to move a whole branch
to a newer "base" commit. This is done by git-rebase. You specify
the branch to move (default HEAD) and where to move it to (no default),
and git cherry-picks every patch out of that branch, applies it on top
of the target, and moves the refs/heads/<branch> pointer to the newly
created commits.
By default, "the branch" is every commit back to the last common
ancestor of the branch head and the target, but you can override that
with command-line arguments.
If you want to avoid merge conflicts due to the master code changing out
from under your edits, but not have "cleanup" merges in your history,
git-rebase is the tool to use.
Git-rebase will also use git-rerere if enabled ("mkdir .git/rr-cache").
If rebasing encounters a conflict it can't resolve, it will stop halfway
and ask you to resolve the problem by hand. However, it still knows it
has a job to finish! The unapplied patches are remembered until you do
one of
git-rebase --continue
This will check in the current index. You should
do git-update-index <files> in the conflicts that
you resolve, but NOT do an actual git-commit.
git-rebase --continue will do the commit.
git-rebase --skip
This will skip the conflicting patch. You
don't have to resolve the conflicts; git will
just back up and try the next patch in the series.
git-rebase --abort
This will abandon the whole rebase operation (including
any half-done work) and return you to where you began.
Git-rebase can also help you divide up work. Suppose you've mixed up
development of two features in the current HEAD, a branch called "dev".
You want to divide them up into "dev1" and "dev2". Assuming that HEAD
is a branch off master, then you can either look through
git log master..HEAD
or just get a raw list of the commits with
git rev-list master..HEAD
Either way, suppose you figure out a list of commits that you want in
dev1 and create that branch:
git checkout -b dev1 master
for i in `cat commit_list`; do
git-cherry-pick $i
done
You can use the other half of the list you edited to generate the dev2
branch, but if you're not sure if you forgot something, or just don't
feel like doing that manual work, then you can use git-rebase to do it
for you...
git checkout -b dev2 dev # Create dev2 branch
git-rebase --onto master dev1 # Subreact dev1 and rebase
This will find all patches that are in dev and not in dev1,
apply them on top of master, and call the result dev2.
* Experimenting with merging
To play with non-trivial merging, get an existing git repository of
a non-trivial project (git itself and the Linux kernel are readily
available. Fire up gitk to look at history, find some interesting-looking
merges, and redo them yourself on a test branch.
As long as you do everything on test branches, you aren't going to screw
anything up. So play!
You can use gitk to search for "Conflicts:" in the commit comments to
find merges that didn't go smoothly and see what happens. (Or you can
search in "git log" output. gitk just draws prettier pictures.)
You can also set up two repositories on the same machine and try pulling
and pushing between them.
To identify arbitrary commits, the 40-byte raw hex ID is probably easiest;
you can cut-and-paste them from the gitk window.
For example, in the git repository,
3f69d405d749742945afd462bff6541604ecd420
looks like an interesting merge. Its parents are
Parent: 7d55561986ffe94ca7ca22dc0a6846f698893226
Parent: 097dc3d8c32f4b85bf9701d5e1de98999ac25c1c
Let's try doing that manually:
$ git checkout -b test 7d55561986ffe94ca7ca22dc0a6846f698893226
$ git pull . 097dc3d8c32f4b85bf9701d5e1de98999ac25c1c
error: no such remote ref refs/heads/097dc3d8c32f4b85bf9701d5e1de98999ac25c1c
Fetch failure: .
Cool! I didn't know that wasn't allowed. (I'll have to ask why it's
not; perhaps it's because it uses the branch name in the automatic
commit message.) I could do it by hand with git-merge, but I'll just
give it a branch name:
$ git branch test2 097dc3d8c32f4b85bf9701d5e1de98999ac25c1c
$ git pull . test2
Merging HEAD with 097dc3d8c32f4b85bf9701d5e1de98999ac25c1c
Merging:
7d55561986ffe94ca7ca22dc0a6846f698893226 Merge branch 'jc/dirwalk-n-cache-tree' into jc/cache-tree
097dc3d8c32f4b85bf9701d5e1de98999ac25c1c Remove "tree->entries" tree-entry list from tree parser
found 2 common ancestor(s):
d9b814cc97f16daac06566a5340121c446136d22 Add builtin "git rm" command
288c0384505e6c25cc1a162242919a0485d50a74 Merge branch 'js/fetchconfig'
Merging:
d9b814cc97f16daac06566a5340121c446136d22 Add builtin "git rm" command
288c0384505e6c25cc1a162242919a0485d50a74 Merge branch 'js/fetchconfig'
found 1 common ancestor(s):
63dffdf03da65ddf1a02c3215ad15ba109189d42 Remove old "git-grep.sh" remnants
Auto-merging Makefile
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in Makefile
Auto-merging builtin.h
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in builtin.h
Auto-merging cache.h
Removing check-ref-format.c
Auto-merging git.c
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in git.c
Auto-merging read-cache.c
Auto-merging update-index.c
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in update-index.c
Renaming apply.c => builtin-apply.c
Auto-merging builtin-apply.c
Renaming read-tree.c => builtin-read-tree.c
Auto-merging builtin-read-tree.c
Auto-merging .gitignore
Auto-merging Makefile
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in Makefile
Auto-merging builtin.h
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in builtin.h
Auto-merging cache.h
Auto-merging fsck-objects.c
Removing git-format-patch.sh
Auto-merging git.c
merge: warning: conflicts during merge
CONFLICT (content): Merge conflict in git.c
Auto-merging update-index.c
Automatic merge failed; fix conflicts and then commit the result.
$ git status
Hey, look, lots of interesting stuff. Particularly, see
# Changed but not updated:
# (use git-update-index to mark for commit)
#
# unmerged: Makefile
# modified: Makefile
# unmerged: builtin.h
# modified: builtin.h
# unmerged: git.c
# modified: git.c
The "unmerged" (a.k.a. "staged") files are ones that need manual resolution.
(I notice that update-index.c isn't listed, despite being mentioned
as a conflict in the message. Can someone explain that?)
Fixing those is easy, but as you can see from the original commit comment
and diffs, there were some additional changes that were necessary to
make that compile.
You can test before committing the change, or do it the git way - commit
anyway, then test and "git commit --amend" with the fixes, of any.
Unlike a centralized VCS, committing is not the same as pushing upstream.
You can use test branches in the repository to save as much work as
you like. While it's still nice to keep the public repository clean,
you don't have to worry about "breaking the tree" every time you commit.
You can do all kinds of stuff in test branches, and clean it up later.
This is why all the git merge tools do the commit without waiting for
you to test it. The merge is usually okay, and it saves time. If not,