[PATCH] Documentation: send/receive.

[git/mingw.git] / Documentation / tutorial.txt

A short git tutorial
====================
May 2005


Introduction
------------

This is trying to be a short tutorial on setting up and using a git
archive, mainly because being hands-on and using explicit examples is
often the best way of explaining what is going on.

In normal life, most people wouldn't use the "core" git programs
directly, but rather script around them to make them more palatable. 
Understanding the core git stuff may help some people get those scripts
done, though, and it may also be instructive in helping people
understand what it is that the higher-level helper scripts are actually
doing. 

The core git is often called "plumbing", with the prettier user
interfaces on top of it called "porcelain".  You may not want to use the
plumbing directly very often, but it can be good to know what the
plumbing does for when the porcelain isn't flushing... 


Creating a git archive
----------------------

Creating a new git archive couldn't be easier: all git archives start
out empty, and the only thing you need to do is find yourself a
subdirectory that you want to use as a working tree - either an empty
one for a totally new project, or an existing working tree that you want
to import into git. 

For our first example, we're going to start a totally new archive from
scratch, with no pre-existing files, and we'll call it "git-tutorial".
To start up, create a subdirectory for it, change into that
subdirectory, and initialize the git infrastructure with "git-init-db":

        mkdir git-tutorial
        cd git-tutorial
        git-init-db 

to which git will reply

        defaulting to local storage area

which is just git's way of saying that you haven't been doing anything
strange, and that it will have created a local .git directory setup for
your new project. You will now have a ".git" directory, and you can
inspect that with "ls". For your new empty project, ls should show you
three entries:

 - a symlink called HEAD, pointing to "refs/heads/master"

   Don't worry about the fact that the file that the HEAD link points to
   doesn't even exist yet - you haven't created the commit that will
   start your HEAD development branch yet.

 - a subdirectory called "objects", which will contain all the git SHA1
   objects of your project. You should never have any real reason to
   look at the objects directly, but you might want to know that these
   objects are what contains all the real _data_ in your repository.

 - a subdirectory called "refs", which contains references to objects.

   In particular, the "refs" subdirectory will contain two other
   subdirectories, named "heads" and "tags" respectively.  They do
   exactly what their names imply: they contain references to any number
   of different "heads" of development (aka "branches"), and to any
   "tags" that you have created to name specific versions of your
   repository. 

   One note: the special "master" head is the default branch, which is
   why the .git/HEAD file was created as a symlink to it even if it
   doesn't yet exist. Basically, the HEAD link is supposed to always
   point to the branch you are working on right now, and you always
   start out expecting to work on the "master" branch.

   However, this is only a convention, and you can name your branches
   anything you want, and don't have to ever even _have_ a "master"
   branch.  A number of the git tools will assume that .git/HEAD is
   valid, though.

   [ Implementation note: an "object" is identified by its 160-bit SHA1
   hash, aka "name", and a reference to an object is always the 40-byte
   hex representation of that SHA1 name. The files in the "refs"
   subdirectory are expected to contain these hex references (usually
   with a final '\n' at the end), and you should thus expect to see a
   number of 41-byte files containing these references in this refs
   subdirectories when you actually start populating your tree ]

You have now created your first git archive. Of course, since it's
empty, that's not very useful, so let's start populating it with data.


        Populating a git archive
        ------------------------

We'll keep this simple and stupid, so we'll start off with populating a
few trivial files just to get a feel for it.

Start off with just creating any random files that you want to maintain
in your git archive. We'll start off with a few bad examples, just to
get a feel for how this works:

        echo "Hello World" > a
        echo "Silly example" > b

you have now created two files in your working directory, but to
actually check in your hard work, you will have to go through two steps:

 - fill in the "cache" aka "index" file with the information about your
   working directory state

 - commit that index file as an object.

The first step is trivial: when you want to tell git about any changes
to your working directory, you use the "git-update-cache" program.  That
program normally just takes a list of filenames you want to update, but
to avoid trivial mistakes, it refuses to add new entries to the cache
(or remove existing ones) unless you explicitly tell it that you're
adding a new entry with the "--add" flag (or removing an entry with the
"--remove") flag. 

So to populate the index with the two files you just created, you can do

        git-update-cache --add a b

and you have now told git to track those two files.

In fact, as you did that, if you now look into your object directory,
you'll notice that git will have added two new objects to the object
store.  If you did exactly the steps above, you should now be able to do

        ls .git/objects/??/*

and see two files:

        .git/objects/55/7db03de997c86a4a028e1ebd3a1ceb225be238 
        .git/objects/f2/4c74a2e500f5ee1332c86b94199f52b1d1d962

which correspond with the object with SHA1 names of 557db... and f24c7..
respectively.

If you want to, you can use "git-cat-file" to look at those objects, but
you'll have to use the object name, not the filename of the object:

        git-cat-file -t 557db03de997c86a4a028e1ebd3a1ceb225be238

where the "-t" tells git-cat-file to tell you what the "type" of the
object is. Git will tell you that you have a "blob" object (ie just a
regular file), and you can see the contents with

        git-cat-file "blob" 557db03de997c86a4a028e1ebd3a1ceb225be238

which will print out "Hello World".  The object 557db...  is nothing
more than the contents of your file "a". 

[ Digression: don't confuse that object with the file "a" itself.  The
  object is literally just those specific _contents_ of the file, and
  however much you later change the contents in file "a", the object we
  just looked at will never change.  Objects are immutable.  ]

Anyway, as we mentioned previously, you normally never actually take a
look at the objects themselves, and typing long 40-character hex SHA1
names is not something you'd normally want to do.  The above digression
was just to show that "git-update-cache" did something magical, and
actually saved away the contents of your files into the git content
store. 

Updating the cache did something else too: it created a ".git/index"
file.  This is the index that describes your current working tree, and
something you should be very aware of.  Again, you normally never worry
about the index file itself, but you should be aware of the fact that
you have not actually really "checked in" your files into git so far,
you've only _told_ git about them.

However, since git knows about them, you can now start using some of the
most basic git commands to manipulate the files or look at their status. 

In particular, let's not even check in the two files into git yet, we'll
start off by adding another line to "a" first:

        echo "It's a new day for git" >> a

and you can now, since you told git about the previous state of "a", ask
git what has changed in the tree compared to your old index, using the
"git-diff-files" command:

        git-diff-files 

oops.  That wasn't very readable.  It just spit out its own internal
version of a "diff", but that internal version really just tells you
that it has noticed that "a" has been modified, and that the old object
contents it had have been replaced with something else.

To make it readable, we can tell git-diff-files to output the
differences as a patch, using the "-p" flag:

        git-diff-files -p

which will spit out

        diff --git a/a b/a
        --- a/a
        +++ b/a
        @@ -1 +1,2 @@
         Hello World
        +It's a new day for git

ie the diff of the change we caused by adding another line to "a".

In other words, git-diff-files always shows us the difference between
what is recorded in the index, and what is currently in the working
tree. That's very useful.

A common shorthand for "git-diff-files -p" is to just write

        git diff

which will do the same thing. 


        Committing git state
        --------------------

Now, we want to go to the next stage in git, which is to take the files
that git knows about in the index, and commit them as a real tree. We do
that in two phases: creating a "tree" object, and committing that "tree"
object as a "commit" object together with an explanation of what the
tree was all about, along with information of how we came to that state.

Creating a tree object is trivial, and is done with "git-write-tree". 
There are no options or other input: git-write-tree will take the
current index state, and write an object that describes that whole
index.  In other words, we're now tying together all the different
filenames with their contents (and their permissions), and we're
creating the equivalent of a git "directory" object:

        git-write-tree

and this will just output the name of the resulting tree, in this case
(if you have does exactly as I've described) it should be

        3ede4ed7e895432c0a247f09d71a76db53bd0fa4

which is another incomprehensible object name. Again, if you want to,
you can use "git-cat-file -t 3ede4.." to see that this time the object
is not a "blob" object, but a "tree" object (you can also use
git-cat-file to actually output the raw object contents, but you'll see
mainly a binary mess, so that's less interesting).

However - normally you'd never use "git-write-tree" on its own, because
normally you always commit a tree into a commit object using the
"git-commit-tree" command. In fact, it's easier to not actually use
git-write-tree on its own at all, but to just pass its result in as an
argument to "git-commit-tree".

"git-commit-tree" normally takes several arguments - it wants to know
what the _parent_ of a commit was, but since this is the first commit
ever in this new archive, and it has no parents, we only need to pass in
the tree ID. However, git-commit-tree also wants to get a commit message
on its standard input, and it will write out the resulting ID for the
commit to its standard output.

And this is where we start using the .git/HEAD file. The HEAD file is
supposed to contain the reference to the top-of-tree, and since that's
exactly what git-commit-tree spits out, we can do this all with a simple
shell pipeline:

        echo "Initial commit" | git-commit-tree $(git-write-tree) > .git/HEAD

which will say:

        Committing initial tree 3ede4ed7e895432c0a247f09d71a76db53bd0fa4

just to warn you about the fact that it created a totally new commit
that is not related to anything else. Normally you do this only _once_
for a project ever, and all later commits will be parented on top of an
earlier commit, and you'll never see this "Committing initial tree"
message ever again.

Again, normally you'd never actually do this by hand.  There is a
helpful script called "git commit" that will do all of this for you. So
you could have just writtten

        git commit

instead, and it would have done the above magic scripting for you.


        Making a change
        ---------------

Remember how we did the "git-update-cache" on file "a" and then we
changed "a" afterward, and could compare the new state of "a" with the
state we saved in the index file? 

Further, remember how I said that "git-write-tree" writes the contents
of the _index_ file to the tree, and thus what we just committed was in
fact the _original_ contents of the file "a", not the new ones. We did
that on purpose, to show the difference between the index state, and the
state in the working directory, and how they don't have to match, even
when we commit things.

As before, if we do "git-diff-files -p" in our git-tutorial project,
we'll still see the same difference we saw last time: the index file
hasn't changed by the act of committing anything.  However, now that we
have committed something, we can also learn to use a new command:
"git-diff-cache".

Unlike "git-diff-files", which showed the difference between the index
file and the working directory, "git-diff-cache" shows the differences
between a committed _tree_ and either the the index file or the working
directory.  In other words, git-diff-cache wants a tree to be diffed
against, and before we did the commit, we couldn't do that, because we
didn't have anything to diff against. 

But now we can do 

        git-diff-cache -p HEAD

(where "-p" has the same meaning as it did in git-diff-files), and it
will show us the same difference, but for a totally different reason. 
Now we're comparing the working directory not against the index file,
but against the tree we just wrote.  It just so happens that those two
are obviously the same, so we get the same result.

Again, because this is a common operation, you can also just shorthand
it with

        git diff HEAD

which ends up doing the above for you.

In other words, "git-diff-cache" normally compares a tree against the
working directory, but when given the "--cached" flag, it is told to
instead compare against just the index cache contents, and ignore the
current working directory state entirely.  Since we just wrote the index
file to HEAD, doing "git-diff-cache --cached -p HEAD" should thus return
an empty set of differences, and that's exactly what it does. 

[ Digression: "git-diff-cache" really always uses the index for its
  comparisons, and saying that it compares a tree against the working
  directory is thus not strictly accurate. In particular, the list of
  files to compare (the "meta-data") _always_ comes from the index file,
  regardless of whether the --cached flag is used or not. The --cached
  flag really only determines whether the file _contents_ to be compared
  come from the working directory or not.

  This is not hard to understand, as soon as you realize that git simply
  never knows (or cares) about files that it is not told about
  explicitly. Git will never go _looking_ for files to compare, it
  expects you to tell it what the files are, and that's what the index
  is there for.  ]

However, our next step is to commit the _change_ we did, and again, to
understand what's going on, keep in mind the difference between "working
directory contents", "index file" and "committed tree".  We have changes
in the working directory that we want to commit, and we always have to
work through the index file, so the first thing we need to do is to
update the index cache:

        git-update-cache a

(note how we didn't need the "--add" flag this time, since git knew
about the file already).

Note what happens to the different git-diff-xxx versions here.  After
we've updated "a" in the index, "git-diff-files -p" now shows no
differences, but "git-diff-cache -p HEAD" still _does_ show that the
current state is different from the state we committed.  In fact, now
"git-diff-cache" shows the same difference whether we use the "--cached"
flag or not, since now the index is coherent with the working directory. 

Now, since we've updated "a" in the index, we can commit the new
version.  We could do it by writing the tree by hand again, and
committing the tree (this time we'd have to use the "-p HEAD" flag to
tell commit that the HEAD was the _parent_ of the new commit, and that
this wasn't an initial commit any more), but you've done that once
already, so let's just use the helpful script this time:

        git commit

which starts an editor for you to write the commit message and tells you
a bit about what you're doing. 

Write whatever message you want, and all the lines that start with '#'
will be pruned out, and the rest will be used as the commit message for
the change. If you decide you don't want to commit anything after all at
this point (you can continue to edit things and update the cache), you
can just leave an empty message. Otherwise git-commit-script will commit
the change for you.

You've now made your first real git commit. And if you're interested in
looking at what git-commit-script really does, feel free to investigate:
it's a few very simple shell scripts to generate the helpful (?) commit
message headers, and a few one-liners that actually do the commit itself.


        Checking it out
        ---------------

While creating changes is useful, it's even more useful if you can tell
later what changed.  The most useful command for this is another of the
"diff" family, namely "git-diff-tree". 

git-diff-tree can be given two arbitrary trees, and it will tell you the
differences between them. Perhaps even more commonly, though, you can
give it just a single commit object, and it will figure out the parent
of that commit itself, and show the difference directly. Thus, to get
the same diff that we've already seen several times, we can now do

        git-diff-tree -p HEAD

(again, "-p" means to show the difference as a human-readable patch),
and it will show what the last commit (in HEAD) actually changed.

More interestingly, you can also give git-diff-tree the "-v" flag, which
tells it to also show the commit message and author and date of the
commit, and you can tell it to show a whole series of diffs.
Alternatively, you can tell it to be "silent", and not show the diffs at
all, but just show the actual commit message.

In fact, together with the "git-rev-list" program (which generates a
list of revisions), git-diff-tree ends up being a veritable fount of
changes. A trivial (but very useful) script called "git-whatchanged" is
included with git which does exactly this, and shows a log of recent
activity.

To see the whole history of our pitiful little git-tutorial project, you
can do

        git log

which shows just the log messages, or if we want to see the log together
with the associated patches use the more complex (and much more
powerful)

        git-whatchanged -p --root

and you will see exactly what has changed in the repository over its
short history. 

[ Side note: the "--root" flag is a flag to git-diff-tree to tell it to
  show the initial aka "root" commit too.  Normally you'd probably not
  want to see the initial import diff, but since the tutorial project
  was started from scratch and is so small, we use it to make the result
  a bit more interesting ]

With that, you should now be having some inkling of what git does, and
can explore on your own.


        Copying archives
        -----------------

Git archives are normally totally self-sufficient, and it's worth noting
that unlike CVS, for example, there is no separate notion of
"repository" and "working tree".  A git repository normally _is_ the
working tree, with the local git information hidden in the ".git"
subdirectory.  There is nothing else.  What you see is what you got. 

[ Side note: you can tell git to split the git internal information from
  the directory that it tracks, but we'll ignore that for now: it's not
  how normal projects work, and it's really only meant for special uses.
  So the mental model of "the git information is always tied directly to
  the working directory that it describes" may not be technically 100%
  accurate, but it's a good model for all normal use ]

This has two implications: 

 - if you grow bored with the tutorial archive you created (or you've
   made a mistake and want to start all over), you can just do simple

        rm -rf git-tutorial

   and it will be gone. There's no external repository, and there's no
   history outside of the project you created.

 - if you want to move or duplicate a git archive, you can do so. There
   is no "git clone" command: if you want to create a copy of your
   archive (with all the full history that went along with it), you can
   do so with a regular "cp -a git-tutorial new-git-tutorial".

   Note that when you've moved or copied a git archive, your git index
   file (which caches various information, notably some of the "stat"
   information for the files involved) will likely need to be refreshed.
   So after you do a "cp -a" to create a new copy, you'll want to do

        git-update-cache --refresh

   to make sure that the index file is up-to-date in the new one. 

Note that the second point is true even across machines.  You can
duplicate a remote git archive with _any_ regular copy mechanism, be it
"scp", "rsync" or "wget". 

When copying a remote repository, you'll want to at a minimum update the
index cache when you do this, and especially with other peoples
repositories you often want to make sure that the index cache is in some
known state (you don't know _what_ they've done and not yet checked in),
so usually you'll precede the "git-update-cache" with a

        git-read-tree --reset HEAD
        git-update-cache --refresh

which will force a total index re-build from the tree pointed to by HEAD
(it resets the index contents to HEAD, and then the git-update-cache
makes sure to match up all index entries with the checked-out files). 

The above can also be written as simply

        git reset

and in fact a lot of the common git command combinations can be scripted
with the "git xyz" interfaces, and you can learn things by just looking
at what the git-*-script scripts do ("git reset" is the above two lines
implemented in "git-reset-script", but some things like "git status" and
"git commit" are slightly more complex scripts around the basic git
commands). 

NOTE! Many (most?) public remote repositories will not contain any of
the checked out files or even an index file, and will _only_ contain the
actual core git files.  Such a repository usually doesn't even have the
".git" subdirectory, but has all the git files directly in the
repository. 

To create your own local live copy of such a "raw" git repository, you'd
first create your own subdirectory for the project, and then copy the
raw repository contents into the ".git" directory. For example, to
create your own copy of the git repository, you'd do the following

        mkdir my-git
        cd my-git
        rsync -rL rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/git.git/ .git

followed by 

        git-read-tree HEAD

to populate the index. However, now you have populated the index, and
you have all the git internal files, but you will notice that you don't
actually have any of the _working_directory_ files to work on. To get
those, you'd check them out with

        git-checkout-cache -u -a

where the "-u" flag means that you want the checkout to keep the index
up-to-date (so that you don't have to refresh it afterward), and the
"-a" file means "check out all files" (if you have a stale copy or an
older version of a checked out tree you may also need to add the "-f"
file first, to tell git-checkout-cache to _force_ overwriting of any old
files). 

Again, this can all be simplified with

        git clone rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/git.git/ my-git
        cd my-git
        git checkout

which will end up doing all of the above for you.

You have now successfully copied somebody else's (mine) remote
repository, and checked it out. 


        Creating a new branch
        ---------------------

Branches in git are really nothing more than pointers into the git
object space from within the ",git/refs/" subdirectory, and as we
already discussed, the HEAD branch is nothing but a symlink to one of
these object pointers. 

You can at any time create a new branch by just picking an arbitrary
point in the project history, and just writing the SHA1 name of that
object into a file under .git/refs/heads/.  You can use any filename you
want (and indeed, subdirectories), but the convention is that the
"normal" branch is called "master".  That's just a convention, though,
and nothing enforces it. 

To show that as an example, let's go back to the git-tutorial archive we
used earlier, and create a branch in it.  You literally do that by just
creating a new SHA1 reference file, and switch to it by just making the
HEAD pointer point to it:

        cat .git/HEAD > .git/refs/heads/mybranch
        ln -sf refs/heads/mybranch .git/HEAD

and you're done.

Now, if you make the decision to start your new branch at some other
point in the history than the current HEAD, you usually also want to
actually switch the contents of your working directory to that point
when you switch the head, and "git checkout" will do that for you:
instead of switching the branch by hand with "ln -sf", you can just do

        git checkout mybranch

which will basically "jump" to the branch specified, update your working
directory to that state, and also make it become the new default HEAD. 

You can always just jump back to your original "master" branch by doing

        git checkout master

and if you forget which branch you happen to be on, a simple

        ls -l .git/HEAD

will tell you where it's pointing.


        Merging two branches
        --------------------

One of the ideas of having a branch is that you do some (possibly
experimental) work in it, and eventually merge it back to the main
branch.  So assuming you created the above "mybranch" that started out
being the same as the original "master" branch, let's make sure we're in
that branch, and do some work there.

        git checkout mybranch
        echo "Work, work, work" >> a
        git commit a

Here, we just added another line to "a", and we used a shorthand for
both going a "git-update-cache a" and "git commit" by just giving the
filename directly to "git commit". 

Now, to make it a bit more interesting, let's assume that somebody else
does some work in the original branch, and simulate that by going back
to the master branch, and editing the same file differently there:

        git checkout master

Here, take a moment to look at the contents of "a", and notice how they
don't contain the work we just did in "mybranch" - because that work
hasn't happened in the "master" branch at all. Then do

        echo "Play, play, play" >> a
        echo "Lots of fun" >> b
        git commit a b

since the master branch is obviously in a much better mood.

Now, you've got two branches, and you decide that you want to merge the
work done. Before we do that, let's introduce a cool graphical tool that
helps you view what's going on:

        gitk --all

will show you graphically both of your branches (that's what the "--all"
means: normally it will just show you your current HEAD) and their
histories.  You can also see exactly how they came to be from a common
source. 

Anyway, let's exit gitk (^Q or the File menu), and decide that we want
to merge the work we did on the "mybranch" branch into the "master"
branch (which is currently our HEAD too).  To do that, there's a nice
script called "git resolve", which wants to know which branches you want
to resolve and what the merge is all about:

        git resolve HEAD mybranch "Merge work in mybranch"

where the third argument is going to be used as the commit message if
the merge can be resolved automatically.

Now, in this case we've intentionally created a situation where the
merge will need to be fixed up by hand, though, so git will do as much
of it as it can automatically (which in this case is just merge the "b"
file, which had no differences in the "mybranch" branch), and say:

        Simple merge failed, trying Automatic merge
        Auto-merging a.
        merge: warning: conflicts during merge
        ERROR: Merge conflict in a.
        fatal: merge program failed
        Automatic merge failed, fix up by hand

which is way too verbose, but it basically tells you that it failed the
really trivial merge ("Simple merge") and did an "Automatic merge"
instead, but that too failed due to conflicts in "a".

Not to worry. It left the (trivial) conflict in "a" in the same form you
should already be well used to if you've ever used CVS, so let's just
open "a" in our editor (whatever that may be), and fix it up somehow.
I'd suggest just making it so that "a" contains all four lines:

        Hello World
        It's a new day for git
        Play, play, play
        Work, work, work

and once you're happy with your manual merge, just do a

        git commit a

which will very loudly warn you that you're now committing a merge
(which is correct, so never mind), and you can write a small merge
message about your adventures in git-merge-land. 

After you're done, start up "gitk --all" to see graphically what the
history looks like.  Notive that "mybranch" still exists, and you can
switch to it, and continue to work with it if you want to.  The
"mybranch" branch will not contain the merge, but next time you merge it
from the "master" branch, git will know how you merged it, so you'll not
have to do _that_ merge again.


        Merging external work
        ---------------------

It's usually much more common that you merge with somebody else than
merging with your own branches, so it's worth pointing out that git
makes that very easy too, and in fact, it's not that different from
doing a "git resolve".  In fact, a remote merge ends up being nothing
more than "fetch the work from a remote repository into a temporary tag"
followed by a "git resolve". 

It's such a common thing to do that it's called "git pull", and you can
simply do

        git pull <remote-repository>

and optionally give a branch-name for the remote end as a second
argument.

[ Todo: fill in real examples ]


        Tagging a version
        -----------------

In git, there's two kinds of tags, a "light" one, and a "signed tag".

A "light" tag is technically nothing more than a branch, except we put
it in the ".git/refs/tags/" subdirectory instead of calling it a "head".
So the simplest form of tag involves nothing more than

        cat .git/HEAD > .git/refs/tags/my-first-tag

after which point you can use this symbolic name for that particular
state. You can, for example, do

        git diff my-first-tag

to diff your current state against that tag (which at this point will
obviously be an empty diff, but if you continue to develop and commit
stuff, you can use your tag as a "anchor-point" to see what has changed
since you tagged it.

A "signed tag" is actually a real git object, and contains not only a
pointer to the state you want to tag, but also a small tag name and
message, along with a PGP signature that says that yes, you really did
that tag. You create these signed tags with

        git tag <tagname>

which will sign the current HEAD (but you can also give it another
argument that specifies the thing to tag, ie you could have tagged the
current "mybranch" point by using "git tag <tagname> mybranch").

You normally only do signed tags for major releases or things
like that, while the light-weight tags are useful for any marking you
want to do - any time you decide that you want to remember a certain
point, just create a private tag for it, and you have a nice symbolic
name for the state at that point.

[ to be continued.. cvsimports, pushing and pulling ]