Git: Difference between revisions

From miki
Jump to navigation Jump to search
Line 395: Line 395:
git config --add daemon.receivepack true # no hyphen in receivepack!
git config --add daemon.receivepack true # no hyphen in receivepack!
</source>
</source>

=== CR/LF conversion ===

* To disable CR/LF completely, across all copies of the repositories, create a file {{file|.gitattributes}} in the root of the repository with the following as content [http://stackoverflow.com/questions/9933004/best-way-to-disable-git-end-of-line-normalization-crlf-to-lf-across-all-clones]:
* -text
:This should no other side-effect than disabling CRLF/LF conversion, and requires no further configuration, including on new/old clones.


== How-To ==
== How-To ==

Revision as of 09:11, 28 July 2014

References

Git cheat sheet

Things to look at

Git vs ...

Git Usage Statistics

According to recent statistics, Git is growing faster than other DVCS system, and would be 10 times more used than e.g. Mercurial:

Git on Windows

Tools

Git SVN

Subversion integration

See Git SVN page

Introduction

Git Features:

  • Reliability
  • Performance
  • Distributed

Distributed

Originally from BitKeeper. Other distributed SCM is Mercurial.

  • No single repository. Everybody always has his own copy of the repository. The repository content is pulled from other people's repository.
  • No politics, no commit access control. All work is always done locally, so there is no need to define such politics.

Reliability

Every change, file, directory, etc. is cryptographically hashed (sha1sum).

  • Easy corruption detection. Any tampering to a file or directory content (either malicious or because of hardware failure) is immediately detected.
  • Easy distribution. Moreover because the repository is distributed all over the place, it is very easy to repair a given repository. You only need to drop all broken objects, and get all missing objects from a remote copy.

Performance

Very fast commit. Local repository

Terminology and Concepts

gitrevisions
A commit is a snapshot of your working tree at some point in time. A revision in git denotes either a given commit, or all commit that can be reached from that commit. There are different ways to name a commit (see also man gitrevisions):
  • branchname — a branch name is an alias for most recent commit on that branch
  • tagname — similar to a branch alias, but that does not change in time
  • HEAD — currently checked out commit
  • c82a22c — the SHA-1 hash id of the commit (can be truncated as long as it remains unique)
  • name^ — the parent of commit name
  • name^^ — the grand-parent of commit name (and so on)
  • name^2 — the 2nd parent of commit name (and so on)
  • name~10 — the 10th ancestor of commit name (same as name^^^^^^^^^^)
  • name:path — reference a specific file/directory in a given commit
  • name^{tree} — reference the tree held by a commit
  • name1..name2 — a commit range, i.e. all commits reachable from name2 back to, but no including, name1 (if either name is omitted, use HEAD instead)
  • name1...name2 — refers to all commits referenced by name1 or name2, but not by both. For git diff, refers to all commits between name2 and the common ancestor of name1 and name2.
  • master.. — to review changes made to the current branch
  • ..master — after a fetch, to review all changes occured since last rebase or merge
  • --since="2 weeks ago" — all commits since a certain date
  • --until=”1 week ago” — all commits up to a certain date
  • --grep=pattern — all commits whose commit message matches the regular expression pattern.
  • --committer=pattern — all commits whose committer matches the pattern
  • --author=pattern — all commits whose author matches the pattern
  • --no-merges — all commits in a range that have only one pattern (i.e. ignore all merge commits)
detached head
When HEAD is no longer a reference to anything (like ref: refs/heads/branch), but instead contains the actual hash of a commit.
git checkout -b newbranch           # To attach HEAD back on a new branch...
hunk
individual change within a file (basically a file diff output is made of a sequence of one or more hunks).

Install

From packages

Install the following essential packages:

  • git-core — the main program
  • git-gui — a gui front-end

Optionally install also:

  • git-doc — documentation
  • gitweb — Web interface
  • ViewGit — Another web interface
  • gitosis — Project management:
  • tig — a text-mode repository browser interface to git and color pager.
tig                                # launch browser
git show | tig                     # Use as pager. Colorize output of git-show
  • gitview — Git Repository browser
  • gitg — a Git repository browser targeting Gtk+ / GNOME
  • Version delivered with Lucid/Maverick is a very old one. Compile from the sources to get the latest version. Alternatively the repository ppa:pasgui/ppa contains a more recent version.
    To install the repository:
    sudo apt-add-repository ppa:pasgui/ppa
    

    or alternatively, create a file /etc/apt/sources.list.d/pasgui-ppa-lucid.list (change lucid as necessary):

    deb http://ppa.launchpad.net/pasgui/ppa/ubuntu lucid main 
    deb-src http://ppa.launchpad.net/pasgui/ppa/ubuntu lucid main
    

    Then add the apt key:

    sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys F599ACE3
    
  • qgit — A graphical interface to git repositories using QT

From sources

Get the sources with

git clone git://github.com/gitster/git.git          # Use proxygit if behind a proxy
cd git
git checkout v1.7.8                                 # To use given version

Some dependencies must be installed:

sudo apt-get install autoconf zlib1g-dev libcurl4-openssl-dev expat asciidoc     # More packages might be needed
sudo apt-get install docbook2x                                                   # For info pages

For cygwin:

  • Install the packages above (or similar)
  • install package docbook-xml45
  • if needed (see INSTALL):
xmlcatalog --noout --add rewriteURI http://docbook.sourceforge.net/release/xsl/current /usr/share/sgml/docbook/xsl-stylesheets /etc/xml/catalog
xmlcatalog --noout --add rewriteURI http://www.oasis-open.org/docbook/xml/4.5/xsl/current /usr/share/sgml/docbook/xml-dtd-4.5 /etc/xml/catalog
  • Add temporarily an executable script ~/bin/xmlto:
#! /bin/bash
/usr/bin/xmlto --skip-validation "$@"

To build and install (again, see INSTALL):

make configure                                           # See also INSTALL
./configure --prefix=/usr/local                          # In // of existing package installation. /usr/local has precedence
make all doc info                                        # 'info' optional, more dependencies
sudo make install install-doc install-html install-info  # 'install-info' optional

Configuration

References:

Global per-user configuration settings are stored in file ~/.gitconfig

  • Add color to git output for all commmands:
  • git config --global color.ui true
    
  • Define author/email
  • git config --global user.name "Your Name"
    git config --global user.email you@example.com
    
  • Add some frequently used aliases:
  • git config --global alias.st    'status'
    git config --global alias.ci    'commit'
    git config --global alias.co    'checkout'
    git config --global alias.br    'branch'
    git config --global alias.ls    'ls-files'
    git config --global alias.last  'log -1 HEAD'
    # h like history
    git config --global alias.h     'log --oneline --graph --decorate -45'
    git config --global alias.ha    'log --oneline --graph --decorate -45 --all'
    # l like log ;-)
    git config --global alias.l     'log --pretty=tformat:\"%C(yellow)%h %Cblue%an %Cgreen%cr %Cred%d %Creset%s\" --graph -45'
    git config --global alias.la    'log --pretty=tformat:\"%C(yellow)%h %Cblue%an %Cgreen%cr %Cred%d %Creset%s\" --graph -45 --all'
    git config --global alias.dc    'diff --cached'
    git config --global alias.wdiff 'diff --color-words'
    git config --global alias.wshow 'show --color-words'
    
  • Setup global .gitattributes file (for better tokenization with git diff --color-words):
  • git config --global core.attributesfile '~/.gitattributes'
    

    Content of ~/.gitattributes:

    *.c    diff=cpp
    *.cpp  diff=cpp
    
  • Some handy scripts:
    • git-wtf displays the state of your repository in a readable and easy-to-scan format
  • Global ~/.gitignore_global file (see [4]):
    # Global .gitignore file that applies to all git repositories
    #
    # Add this file to git global config with
    #
    #   git config --global core.excludesfile ~/.gitignore_global
    
    # Vim temp files
    .*.swp
    

    Add this file with:

    git config --global core.excludesfile ~/.gitignore_global
    
  • Filter script to prevent running git clean in HOME directory:
    #! /bin/bash
    
    if [ -a /usr/local/bin/git ]; then
      ORIG_GIT=/usr/local/bin/git
    else
      ORIG_GIT=/usr/bin/git
    fi
    
    for c; do
        if [ "$c" = "clean" ]; then
          GIT_HOME="$HOME/.git"
          GIT_DIR="$($ORIG_GIT rev-parse --git-dir 2>/dev/null)"
          GIT_HOME_STAT=$(stat -c %i "$GIT_HOME")
          GIT_DIR_STAT=$(stat -c %i "$GIT_DIR")
          if [ "$GIT_HOME_STAT" = "$GIT_DIR_STAT" -o "$GIT_HOME" = "$GIT_DIR" ]; then
            echo "Can't invoke 'git clean' in HOME git repository!"
            exit 1
          fi
          break                                   # Break on 'clean'
        fi
        if [[ $c != -* || $c == -- ]]; then
            break                                 # ... or on first non-option arg or --
        fi
    done
    
    "$ORIG_GIT" "$@"
    
  • Disable pager for git log:
  • git config --global pager.log false           # Can be enabled back with 'git -l log'
    
  • Enable auto-complete for git aliases
  • alias g=git
    alias gg='git -p'
    complete -F _git g                          # Enable auto-complete for these aliases too
    complete -F _git gg
    

gitk and tcl/tk

  • To solve the issue of ugly fonts (not anti-aliased besides other uglyness), install tk8.5 and force alternatives for wish (see [5]):
  • sudo apt-get install tk8.5
    sudo update-alternatives --config wish
    # select wish8.5
    

    An alternative however is to use gitg.

  • To use Gnome GTK native look, install tileqt as tcl/qt theme (see [6], instructions from [7] ):
  • # Install tileqt build deps
    sudo apt-get install tcl8.5-dev tk8.5-dev libqt4-dev checkinstall
    
    
    # Fetch, configure, build tileqt
    cd ~/tmp
    proxygit clone git://tktable.git.sourceforge.net/gitroot/tktable/tile-qt
    cd tile-qt
    # On UBUNTU LUCID:
    ./configure --with-tcl=/usr/lib/tcl8.5/ --with-tk=/usr/lib/tk8.5/
    # On UBUNTU PRECISE (multi-arch), pick the correct arch:
    ./configure --with-tcl=/usr/lib/tcl8.5/ --with-tk=/usr/lib/tk8.5/ --with-qt-include=/usr/include/qt4 --with-qt-lib=/usr/lib/i386-linux-gnu
    ./configure --with-tcl=/usr/lib/tcl8.5/ --with-tk=/usr/lib/tk8.5/ --with-qt-include=/usr/include/qt4 --with-qt-lib=/usr/lib/x86_64-linux-gnu
    make
    sudo checkinstall                                             # Answer questions until files are copied
    
    # Set 'tileqt' as default theme, and update system (w/o reboot)
    echo '*TkTheme: tileqt' >> ~/.Xresources
    xrdb -merge ~/Xresources
    

    More info/themes:

    Issues:

    • Does not work on Ubuntu Precise (at least under VB). We get the message
      X error of failed request: BadDrawable (invalid Pixmap or Window parameter)

Hook - prevent non-fast-forwarded update

References:

Save this file as .git/hooks/update to install the hook. Edit configuration settings hooks.allownonffupdatemaster to allow / forbid (default) non-fast-forward update.

#!/bin/sh
#
# A hook to prevent non fast-forward update of 'master'.
#
# Config
# ------
# hooks.allownonffupdatemaster
#   This boolean sets whether non-fast-forwarded update on master are allowed.
#   By default, they won't be to prevent history loss.

# --- Command line
refname="$1"
oldrev="$2"
newrev="$3"

# --- Safety check
if [ -z "$GIT_DIR" ]; then
	echo "Don't run this script from the command line." >&2
	echo " (if you want, you could supply GIT_DIR then run" >&2
	echo "  $0 <ref> <oldrev> <newrev>)" >&2
	exit 1
fi

if [ -z "$refname" -o -z "$oldrev" -o -z "$newrev" ]; then
	echo "Usage: $0 <ref> <oldrev> <newrev>" >&2
	exit 1
fi

# --- Config
allownonffupdatemaster=$(git config --bool hooks.allownonffupdatemaster)

# echo "update hook: refname '$refname' oldrev '$oldrev' newrev '$newrev' allownonffupdatemaster '$allownonffupdatemaster'"

case "$refname" in
	refs/heads/master)
		if [ "$allownonffupdatemaster" != "true" ]; then
		    # update is not a fast-forward if there are commits that we can access from oldrev, but not from newrev
		    is_non_ff=$(git rev-list $newrev..$oldrev|wc -l)
		    if [ $is_non_ff -ne 0 ]; then
		        echo "*** [hook] Non-fast-forward updates on master are not allowed, even if forced."
		        exit 1
		    fi
		fi
		;;
	*)
		# Anything else allowed
		exit 0
		;;
esac

# --- Finished
exit 0

Install / Configure git-daemon

References:

git-daemon is the daemon that enables access to git repositories through the git: protocol (port 9418). It is easily installed thanks to package git-daemon-sysvinit:

sudo apt-get install git-daemon-sysvinit
sudo vi /etc/default/git-daemon             # Change user to 'git'
sudo /etc/init.d/git-daemon start           # Start daemon

By default you cannot access any git repository through git: protocol. You must explicitly list which repository are accessible:

cd ~git/git/project.git
touch git-daemon-export-ok                                       # This allow git-daemon to export this project
sudo ln -sf ~git/git/project.git /var/cache/git/project.git      # In addition, project must be listed in /var/cache/git

Note that base address for git-daemon is /var/cache. Hence, given example above, one would clone it with git clone git://server/git/project.git.

Write access are also refused by default. To allow them, you must edit the .git/config file:

cd ~git/git/project.git
git config --add daemon.receivepack true    # no hyphen in receivepack!

CR/LF conversion

  • To disable CR/LF completely, across all copies of the repositories, create a file .gitattributes in the root of the repository with the following as content [8]:
* -text
This should no other side-effect than disabling CRLF/LF conversion, and requires no further configuration, including on new/old clones.

How-To

Here we shall describe how to perform some tasks in Git.

Cloning to/from a Server using SSH

Reference: [9], [10], [11]

Standard peer-2-peer clone

The standard method is simply to use the clone command:

git clone user@server:repositories/myproject

This assumes one has SSH access to user@server.

Bare central repo - SSH access to client

If the clone is not meant to be used locally, but only as a central repository that other users will clone from, it is better to create a bare repository:

git clone --bare user@server:repositories/myproject myproject.git

By default, clone adds the source server as remote. This is usually not necessary:

cd myproject.git
git remote remove origin

Another method is to use git fetch in that case (which does not create a remote):

git init --bare myproject.git
cd myproject.git
git fetch user@server:repositories/myproject "*:*"          # Fetch from local repository
Bare central repo - restricted access from client

Usual, the central git server can't access to remote clients, and also, only offers restricted access from client (via SSH git-shell for instance). In that case, the method is to create an empty bare repository on the server (either via console or via some web interface):

ssh git@griffin
cd repositories
git init --bare myproject.git

And then, from the client, add the server as a remote and push all references:

git remote add git@griffin:repositories myproject.git
git push --all origin
Bare central repo - full access from client

In this method, we create a local bare copy, and then copy it to the remote server.

git clone --bare myproject myproject.git          # Create a bare clone of your repository, if not available yet
cd myproject.git
git remote remove origin
cd ..
scp -r myproject.git/ git@griffin:repositories/   # Copy the repository to server - requires full SSH access
rm -rf myproject.git                              # Delete local bare clone

Alternatively, one can transfer the repository using another user, but then access conditions and file ownership must be corrected.

See git-clone below for more details.

Cloning from a Server using SSH (limited access)

Reference: [12]

Say you have an SSH access to a server but git-core is not installed you you can't install it yourself (for instance on a shared hosting server). You can still use git but it requires some "hacking":

  1. First copy the executables from package git-core, directory /usr/bin to some directory on the server where you have write access (say private/bin). Note that somes files are actually symlinks:
  2. private/bin/: -rwxr-xr-x user webusers git* lrwxrwxrwx user webusers git-receive-pack -> git* -rwxr-xr-x user webusers git-shell* lrwxrwxrwx user webusers git-upload-archive -> git* -rwxr-xr-x user webusers git-upload-pack*
  3. Clone from the server using -u command-line switch:
  4. git clone -u </path/to/private/bin/git-upload-pack> user@server:private/git/myproject.git
    
  5. Edit the local myproject/.git/config file to add the lines marked with a +:
    [remote "origin"]
         fetch = +refs/heads/*:refs/remotes/origin/*
         url = user@server:private/git/myproject.git
    +    uploadpack = /path/to/private/bin/git-upload-pack
    +    receivepack = /path/to/private/bin/git-receive-pack
    

Using a remote with custom port

Just use the generic address ssh://username@hostname:port/...

git clone ssh://username@hostname:22222/home/git/repo/myrepo
git remote add origin ssh://username@hostname:22222/home/git/repo/myrepo

Mirroring

Reference: [13]

Cloning from a Server using git: Protocol over a Proxy

Reference: [14], [15]

The referenced links propose some script. Here another variant. Add this script to your path (say ~/bin/proxygit):

#!/bin/bash
# proxygit - git through http proxy
#
# Usage:  proxygit [options] COMMAND [ARGS]
#   Setup Git HTTP proxy variable GIT_PROXY_COMMAND and call git with the given parameters.
#   The proxy settings are read from env variable $http_proxy
#
# Note:
# - Requires package socat
# - $GIT_PROXY_COMMAND must not be defined

if [ -n "$GIT_PROXY_COMMAND" ]; then
	PROXY=$(echo $http_proxy|sed -r 's!(http://)?([^/:]*):([0-9]*)/?!\2!')
	PROXYPORT=$(echo $http_proxy|sed -r 's!(http://)?([^/:]*):([0-9]*)/?!\3!')
	exec /usr/bin/socat - "PROXY:$PROXY:$1:$2,proxyport=$PROXYPORT"
else
	export GIT_PROXY_COMMAND="$0"
	exec git "$@"
fi

Work with local and remote branches

Let's assume you have already a remote repository setup, like you would obtain if you clone a remote repository:

git clone user@server:repositories/project.git

By issuing this command, git will automatically:

  • Create a local repository, clone of the remote one
  • Call the remote repository origin
  • Create a local branch master, and
  • Configure it to track the branch master on origin. That remote branch is called locally origin/master.

The following commands can be used to create, track, delete local and remote branches.

# CREATE a local branch
git branch newbranch                          # Create a new local branch 'newbranch'
                                              # ... use "git checkout newbranch" to check it out

# CREATE & CHECKOUT a local branch
git checkout -b newbranch                     # Create a new local branch 'newbranch' and check it out

# PUBLISH a local branch (for TRACKing)
# git push [remotename] [localbranch]:[remotebranch]
git push origin serverfix                     # Push a local branch 'serverfix' to remote (create it if necessary)
git push -u origin serverfix                  # ... same but also set upstream reference for TRACKING
git push -u origin serverfix:serverfix        # ... same as above
git push -u origin serverfix:coolfix          # ... same but call the branch 'coolfix' on the remote

# TRACK a remote branch
git branch --track sf origin/serverfix        # Create a local branch 'sf' that tracks remote branch 'serverfix'
git branch --set-upstream sf origin/serverfix # ... same, but when local branch 'sf' already exists

# TRACK & CHECKOUT a remote branch
git checkout --track origin/serverfix         # Checkout a new local branch 'serverfix' to track remote branch 'serverfix'
                                              #   (remember that this branch is called locally 'origin/serverfix')
git checkout -b sf origin/serverfix           # ... same as above, but the local branch is named 'sf'

# FETCH / UPDATE from remote
git fetch
git fetch --prune                             # After fetching, remove any remote tracking branches that no longer exist on the remote

# FETCH from & MERGE with remote
git pull
git pull --prune                             # After fetching, remove any remote tracking branches that no longer exist on the remote

# DELETE a branch
git branch -d sf                              # Delete local branch 'sf'
git branch -d -r origin/serverfix             # Delete remote tracking branch 'serverfix'
git push origin :serverfix                    # Delete branch 'serverfix' on 'origin'
                                              # (basically this means push nothing to remote 'serverfix')

In summary:

  • Use git branch to create, update, delete branches on the local repository.
  • Use git checkout to checkout (possibly new) local branches.
  • Use git push to update the remote, possibly publishing or deleting branches.

Define a diff textconv filter

This applies a diff filter, but only for git diff and git log commands:

  • Edit file ~/.gitattributes, add (! no quotes around the diff parameter)
*.adr       diff=bookmarks_adr
  • Edit file ~/.gitconfig, add
[diff "bookmarks_adr"]
    textconv = sed -r '/NAME=|URL=|SHORT NAME=/!d'

Duplicate a work tree at a given commit without duplicating the .git repo

This can be done with git archive, which is a bit like the svn export function of svn:

# create current subtree at version commit in ~/tmp/project/sub
# !!! DON'T FORGET *trailing slash* in --prefix
git archive --prefix project/sub/ commit . | tar -x -C ~/tmp   
# Same as above
git archive --prefix project/ commit sub | tar -x -C ~/tmp

Pushing to a non-bare repository

(see [16])

Let's consider we have

  • 2 repositories,
  • one on a machine work and
  • another one on a machine home,
  • we cannot push to home because it is not visible

By default it is not possible to push to work because it is not a bare repository. However it is possible to push to another branch than the current one:

# On machine home, pushing to branch 'from-home' on work
git push work HEAD:from-home
# On machine work
git merge from-home

We can configure home to always push master to branch from-home on remote work:

git config remote.work.push +master:from-home

Fetch a single branch from a remote

Say you have a big remote repository, and you want to only fetch a single branch from it so that to save disk space. This is easily done as follows:

git remote add <remote> <url>
git fetch <remote> <remotebranch <localbranch>    # This will fetch branch <remotebranch> from <remote> and call it <localbranch> locally

Restore file timestamp

Creation / modification timestamps are not kept by Git, but one can set the file timestamp to e.g. commit timestamp using the following script (from [17]):

for FILE in $(git ls-files)
do
    TIME=$(git log --pretty=format:%cd -n 1 --date=iso $FILE)
    TIME=$(date -j -f '%Y-%m-%d %H:%M:%S %z' "$TIME" +%Y%m%d%H%M.%S)
    touch -m -t $TIME $FILE
done

Commands

Here we'll summarize how to use some of the Git commands

Frequently used & common command options

Option Valid for Description
-p, -u, --patch log Generate patch
-b log, diff, show Ignore whitespace changes
--oneline log, show Show commit # + description on oneline
-M log, diff, show Find renames
--summary log, diff, show Show summary of changes (file creation, rename...)
--name-status log, diff, show List file name and type of change (Add, Copied, Deleted, Modified, Renamed)

git-add

git-add adds file contents to the index

git add -A                                    # Stage all modified AND new files in current directory and recursively

git-archive

git-archive creates an archive of file from a named tree

# Duplicate current subtree at version commit in ~/tmp/project/sub
# !!! DON'T FORGET *trailing slash* in --prefix
git archive --prefix project/sub/ commit . | tar -x -C ~/tmp   
# Same as above
git archive --prefix project/ commit sub | tar -x -C ~/tmp

git-blame

git blame -w file                                       # Blame file, ignore whitespaces
git blame file | sort | uniq -w 8 | egrep -o "^.{8}"    # Show list of commits that produced file
                                                        # (! actually a subset of git log)

git-branch

git-branch lists, creates, or deletes branches

# CREATE a local branch
git branch newbranch                          # Create a new local branch 'newbranch'
                                              # ... use "git checkout newbranch" to check it out
 
# TRACK a remote branch
git branch --track sf origin/serverfix        # Create a local branch 'sf' that tracks remote branch 'serverfix'
git branch --set-upstream sf origin/serverfix # ... same, but when local branch 'sf' already exists
 
# DELETE a branch
git branch -d sf                              # Delete local branch 'sf'
git branch -d -r origin/serverfix             # Delete remote tracking branch 'serverfix' (see remark below)
git remote prune origin                       # Prune all state remote tracking branch

# MOVE a branch
git branch -f branch commit                   # Move tip of an existing branch to a different commit

# VIEW branches
git branch --all                              # List ALL branches
git branch --contains commit                  # List LOCAL branches that CONTAIN commit
git branch --all --contains commit            # List ALL branches that CONTAIN commit

Note:

  • You can also track local branch (git branch -t local1 local2), but is it useful?
  • Deleting remote tracking branch (git branch -d -r) on the local repository only makes sense if the remote branch has been deleted on the remote, or if git-fetch has been configured not to import that branch anymore. So the best is simply to prune remote tracking branches automatically:
  • git remote prune origin       # Remove all remote tracking branches that no longer exist on the remote (i.e. stale branches)
    git fetch --prune             # After fetching, remove all remote tracking branches that no longer exist on the remote
    

git-checkout

git-checkout checkouts a branch or paths to the working tree.

# CHECKOUT a local branch
git checkout mybranch                         # Checkout an existing branch

# CREATE & CHECKOUT a local branch
git checkout -b newbranch                     # Create a new local branch 'newbranch' and check it out
 
# TRACK & CHECKOUT a remote branch
git checkout --track origin/serverfix         # Checkout a new local branch 'serverfix' to track remote branch 'serverfix'
                                              #   (remember that this branch is called locally 'origin/serverfix')
git checkout -b sf origin/serverfix           # ... same as above, but the local branch is named 'sf'

# Checkout <paths> from INDEX or COMMIT
git checkout -- <paths...>                    # DISCARD changes
git checkout COMMIT -- <paths...>             # Same as "git reset commit -- path; git checkout -- path"

git-clean

git-clean removes untracked files from the working tree.

git clean -dfx                                # force + delete dirs + delete also ignored files

git-clone

git-clone is mainly used to create a local copy of a remote repository, or to create a bare repository (i.e. one without a working tree) for remote storage:

  • Clone a remote repository:
  • git clone git@griffin:repositories/myproject.git  # Clone repository and create working tree in myproject/
    
  • Create a bare repository for remote storage:
  • git clone --bare myproject myproject.git          # Create a bare clone of your repository, if not available yet
    scp -r myproject.git/ git@griffin:repositories/   # Copy the repository to server - requires SSH access
    rm -rf myproject.git                              # Delete local bare clone
    


The command git clone /dir/repo/project.git is identical to running the following commands:

git init                                            # Create an empty repo
git remote add -f origin /dir/repo/project.git      # Add a remote repo called 'origin' and fetch
git set-head -a                                     # Set default remote branch for remote 'origin' automatically
git checkout --track origin/master                  # Create a tracking branch 'master', and update working tree

Another equivalent option for last command is git checkout -b origin origin/master (since the start-point is remote, git creates a tracking branch).

Some variants:

  • To get a local copy of a remote repository, but without changing the working tree (i.e. keeping all local changes), just change the last command to:
git init
git remote add -f origin -m master /dir/repo/project.git
git branch --track master origin/master             # Branch 'master' set up to track remote branch 'master' from 'origin'
  • To merge remote branch locally, but without creating a tracking branch, change the last command to :
# ...
git merge origin/master                             # Merge

git-commit

git commit -m "commit message"    # Gives immediately the commit message on the command-line
git commit -a                     # Add all changes and commit in one pass
git commit --amend                # Amend tip current branch (message, add some files) - also for merge commits
git commit -C <commit>            # Reuse commit message from <commit>

Add the following line to your file ~/.gitconfig:

git config --global alias.ci 'commit'

Now you can use ci instead of commit:

git ci -m "commit message"

git-diff

git-diff shows changes between commits, commit and working tree, etc

Sample of useful options (some can be combined):

git diff --stat <commit>                # Show changes summary as a graph (with + and -)
git diff -b <commit>                    # Ignore whitespaces
git diff -w <commit>                    # Ignore all whitespaces (remove them completely)
git diff --name-only <commit>           # Only list names of changed files
git diff --relative[=<dir>] <commit>    # Restrict changes to current (or given) directory and show relative pathname
git diff -M --summary <commit>          # Show summary of changes (incl. file creation), find renames

git-fetch

git-fetch downloads objects and refs from another repository.

git fetch -p                      # After fetching, remove any remote-tracking branches which no longer exist on the remote
                                  # (see also git remote prune origin)

git-filter-branch

git-filter-branch rewrites branches.

This command can be used for instance to rewrite / rebase a branch while changing the author/committer name/email (see [18] and [19]):

# Set the committer name to author name, for instance after a rebase
git filter-branch --commit-filter '
    export GIT_COMMITTER_NAME="$GIT_AUTHOR_NAME"; 
    export GIT_COMMITTER_EMAIL="$GIT_AUTHOR_EMAIL"; 
    export GIT_COMMITTER_DATE="$GIT_AUTHOR_DATE"; 
    git commit-tree "$@"' -- basecommit..HEAD
# Change the author/committer name for specific commit (e.g. if wrong email was used):
git filter-branch --commit-filter '
        if [ "$GIT_COMMITTER_NAME" = "<Old Name>" ];
        then
                GIT_COMMITTER_NAME="<New Name>";
                GIT_AUTHOR_NAME="<New Name>";
                GIT_COMMITTER_EMAIL="<New Email>";
                GIT_AUTHOR_EMAIL="<New Email>";
                git commit-tree "$@";
        else
                git commit-tree "$@";
        fi' HEAD

Note that basecommit..HEAD can easily be changed to other commit specification, like --since="1 year ago"

Another example to remove sensitive files from repository (from [20]):

# Delete the file(s)
git clone git@github.com:defunkt/github-gem.git
cd github-gem/
git filter-branch --index-filter 'git rm --cached --ignore-unmatch pattern1 pattern2 ...' HEAD 
# .... Add --tag-name-filter "cat" to keep tags - will overwrite existing tags!

# Push to origin
git push origin master --force

# Force cleanup (a safer method is to clone the repository and keep the clone if everything's fine)
rm -rf .git/refs/original/
git reflog expire --expire=now --all
git gc --prune=now
git gc --aggressive --prune=now

Deleting a directory in all commits

# -- --all apply to all branches
# filter-branch -f to ignore backup in refs/original/
git filter-branch -f --index-filter 'git rm --ignore-unmatch --cached -rfq shared/saves' -- --all

Moving a directory

# Rewrite revision history - Move directory (since we branched off master)
git filter-branch --tree-filter 'mv subsystem/inc inc' master..HEAD

Applying on all branches

git filter-branch ... HEAD will only filter all commits that are reachable from HEAD. To filter all branches, simply replace HEAD by -- --all (see [21]):

git filter-branch --tree-filter 'mv subsystem/inc inc' -- --all

Move a directory and symlink

Same as above but we want to keep a symlink to the new location. It appears we can't move and symlink at the same time because Git complains about following files through symlinks. So we proceed in 2 steps:

git filter-branch --tree-filter "moveit nfc4java bb/firmware_35g/option/jvm mv" nfcbase..HEAD
git filter-branch --tree-filter "moveit nfc4java bb/firmware_35g/option/jvm ln" nfcbase..HEAD

moveit is a script in the path with the following code:

#! /bin/bash

SRC=$1
DST=$2
PHASE=$3             # either 'mv' (1st phase) or 'ln' (2nd phase)

if [ "$PHASE" = "mv" ]; then
    # 1st phase

    mv $SRC/nfc.h $DST/include/nfc.h
    mv $SRC/nfc.c $DST/src/nfc.c
    if [ -a $SRC/fri-lib ]; then 
        rm -rf $DST/nfc-fri
        mv $SRC/fri-lib $DST/fri-lib
    else
        rm -rf $DST/nfc-fri
        mv $SRC/nfc-fri $DST/nfc-fri
    fi
else
    # 2nd phase

    ln -sf ../$DST/include/nfc.h $SRC/nfc.h
    ln -sf ../$DST/src/nfc.c $SRC/nfc.c
    if [ -a $DST/fri-lib ]; then
        mv $DST/fri-lib $DST/nfc-fri
        ln -sf ../$DST/nfc-fri $SRC/fri-lib
    else
        ln -sf ../$DST/nfc-fri $SRC/nfc-fri
    fi
fi

As a result the file / directory nfc.h, nfc.c, fri-lib, nfc-fri are moved from nfc4java/ to bb/firmware_35g/option/jvm, and symlinks are created in the source directory so these files / directories are still accessible.

Adding a CR/PR tag to each commmit

Say we need to add a CR/PR tag to each of our commit (like [CR_1234]), we can do it with:

git filter-branch -f --msg-filter "sed -r '1 s/$/ [CR_1234]/'" master..HEAD

git-grep

git-grep prints lines matching a pattern (see also Git Book)

# Find all occurences of pattern in all files committed since last year
for i in $(git log --oneline --all --graph --since="1 year ago" | egrep -o " [a-h0-9]{7} "); do git grep pattern $i; done

Maybe a better solution from [22]:

git grep <regexp> $(git rev-list --all)

git-log

git-log shows commit logs.

git log                                                # Standard history log
git log -5                                             # Limit to 5 commits
git log -- file                                        # List commits affecting file
git log -p -- file                                     # History log, show patch/diff for file
git log -p -M -- file                                  #  ... idem, but find also renames
git log --summary -M -- file                           # Condensed summary with renames
git log --stat -1                                      # Show diff-stat for last commit
git log -w master.. --stat -- $(find . -name "*.java") # Show all changes to java files since branching off master
git log -p topic..HEAD -- .                            # Show all changes done in current dir that are not in topic branch
git log --all --ancestry-path A..                      # Show all commits that are descendants of A (also with gitk)

Some handy aliases:

[alias]
    # h like history
    h = log --oneline --graph --decorate -45
    ha = log --oneline --graph --decorate -45 --all
    # l like log ;-)
    l = log --pretty=tformat:\"%C(yellow)%h %Cblue%an %Cgreen%cr %Cred%d %Creset%s\" --graph -45
    la = log --pretty=tformat:\"%C(yellow)%h %Cblue%an %Cgreen%cr %Cred%d %Creset%s\" --graph -45 --all

View history of a single files, across copy and renames:

git l --follow -- file                                  # Show all commits that modified file (incl. renames)
                                                        # (includes commits that deleted lines, unlike blame)
git log -b --follow -p -- file                          # ... same, show patch and ignore blanks

CAUTION!git log will not show any differences for merge commits! To fix that, use options -m, -c or --cc (see [23]):

git log -p <commit>            # SHOW NOTHING!!!
git log -p -m <commit>         # Show 2 diffs, one for each parent
git log -p -c <commit>         # ... Same as above, but simultaneously
git log -p --cc <commit>       # ... Same as above, but remove hunks that are identical to one of the parent

CAUTION!git log --oneline --graph output can be misleading! Overlooking the graph may give the impression that a given commit is in some branch where actually it is in another branch. Look at the asterisks (*) carefully!. Example:

* | 2dbfa34 frq05215 4 weeks ago Integration implementation...
* |   cd1d5a7 beq03416 4 weeks ago Merged head of trunk into...
|\ \  
* | | 84fb520 frq05215 4 weeks ago Integration implementation...  # This commit is not in same branch as 0eef5c7!
| | | * 0eef5c7 beq03416 4 weeks ago relaxed the rate adaptation...
| | |/  
| | * b998ebd nxp12661 4 weeks ago [PT_TLMT2086] Changed bool_t type into J9bool_t...

git-pull

git-pull fetches from and merges with another repository or a local branch.

git pull -p                   # After fetching, remove any remote-tracking branches which no longer exist on the remote
                              # (see also git remote prune)

git-push

git push updates remote refs along with associated objects. In layman english, git push basically pushes changes to the remote repository, possibly creating, updating, deleting branches, objects or references.

# PUBLISH a local branch (for TRACKing)
# git push [remotename] [localbranch]:[remotebranch]
git push                                      # Push current branch to remote
git push origin                               # Push all matching branches to remote
git push origin serverfix                     # Push a local branch 'serverfix' to remote (create it if necessary)
git push -u origin serverfix                  # ... same but also set upstream reference for TRACKING
git push -u origin serverfix:serverfix        # ... same as above
git push -u origin serverfix:coolfix          # ... same but call the branch 'coolfix' on the remote

# DELETE a branch
git push origin :serverfix                    # Delete remote branch 'serverfix' on 'origin'
                                              # (basically this means push nothing to remote branch 'serverfix')

git-rebase

git-rebase forward-port local commits to the updated upstream head.

git rebase master                                  # Rebase current branch to master branch
git rebase master topic                            # Rebase branch topic to branch master. Same as 
                                                   #   git checkout topic
                                                   #   git rebase master
git rebase master topic --onto newbase             # Same as above, but instead write the commits at dest
git rebase --committer-date-is-author-date master  # ... keep original (author) date for new commits

git-reflog

git reflog manages reflog information. This is very handy to repair mistakes, or to recover lost commits (e.g. after modifying the head of a branch that was the only reference to a given commit)

git reflog

git-remote

Add another remote to an existing repository. Using git remote set-head and git branch one call also set up automatically origin/HEAD and set tracking branches. For instance the following is equivalent to what's done by git clone:

git remote add -f origin git@griffin:repositories/my_project.git  # Add remote repository and fetch automatically
git remote set-head -a origin                                     # Set origin/HEAD automatically - see man git-remote, set-head
git branch --set-upstream master origin                           # Set master to track head (here origin/master) branch from origin
git-remote set-head
Sets / deletes the default branch for a remote. For remote origin, this creates the reference refs/remotes/origin/HEAD with content ref: refs/remotes/origin/master if default branch is master.
Use set-head to follow the changes in another branch than the default one:
git remote set-head origin -a                       # Set default remote branch for remote 'origin' automatically
git diff origin                                     # -> Will show difference with origin/master (if 'master' is the default)
git remote set-head origin exoticbranch             # ... or set it to a different branch (here 'exoticbranch')
git diff origin                                     # -> Now will show diff with origin/exoticbranch

git-reset

git reset has two purposes:

  • commit paths...Index
    Copy paths... from HEAD (or given commit) to index (working tree left unchanged), or
  • branch head → commit
    Move current branch head to HEAD (or given commit), and optionally update index and/or working tree.
  • Modes are:
    mode index working tree preserve local changes
    --soft Yes
    --mixed (default) reset Yes
    --hard reset reset No
    --merge reset keep if idx = commit, else
    reset if idx = tree, else
    abort
    No
    (fully staged changes are lost)
    --keep reset keep if HEAD = commit, else
    reset if HEAD = tree, else
    abort
    Yes

Note:

  • git reset copies the old head to ORIG_HEAD.
  • Use git checkout <commit> -- <paths...> to copy paths... from a given commit to the index and working tree in one go.
  • Caution! — Don't move the head of a branch if changes have been already published, or this will create many merge conflicts.
  • Caution!git reset --hard is a dangerous command. Prefer one of its alternatives, like:
Commit any changes first
git commit -a -m "snapshot WIP"
git reset --hard~3
Stash the changes
git stash
git reset --hard HEAD~3
# ...
git reset --hard HEAD@{1} # or ORIG_HEAD
git stash apply
Idem & don't change master too early
git stash
git checkout -b new-branch HEAD~3
...
git branch -D master
git branch -m new-branch master



Some use cases (see man git-reset for details):

  • Undo a commit and redo
git commit ...
git reset --soft HEAD^
edit
git commit -a -c ORIG_HEAD # or -C
  • Undo commits permanently
git commit ...
git reset --hard HEAD~3
  • Undo a commit, making it a topic branch
git branch topic/wip
git reset --hard HEAD~3 # or --mixed
git checkout topic/wip  # or -m topic/wip
  • Undo a merge or pull
git pull                # conflicts
git reset --hard
git pull . topic/branch # no conflict
git reset --hard ORIG_HEAD
  • Undo a merge or pull inside a dirty work tree
git pull
git reset --merge ORIG_HEAD
  • Interrupted workflow
git checkout feature
#work work work
git commit -a -m "snapshot WIP"
git checkout master
#fix fix fix
git commit
git checkout feature
git reset HEAD^     # or --soft
  • Reset a single file in the index
git add foo.c
git reset -- foo.c

git-show

git-show shows various types of objects.

git show --stat <commit>         # Show commit; for merge, show *combined* commit
git show -b -R <commit>          # Show reverse patch, ignoring blanks (handy for reversing commit by hand)
git show <treeish>:<file>        # Show file at given commit (see also 'git checkout')
git show -s <commit>             # Show commit, but suppress diff output (like 'git log -1')

Tips

Frequently Used Commands

git commit -a                     # Add all changes and commit in one pass
git commit --amend                # Amend tip current branch (message, add some files) - also for merge commits

Working the Git Way

  • Check project diff before commit -a:
  • git diff                          # First see what's in the working tree (or git status)
    git commit -a                     # Commit all changes
    
  • Give git commit a directory argument instead of using -a:
  • git commit fs/                     # Commit all changes in directory fs
    
  • Clean up an ugly sequence of commits ([24]).
  • Better than hunk-based commit because (1) each stage can be tested individually, (2) intermediate commits may contain changes that is not in the final one.
    1. First make sure that the ugly sequence is on some temporary branch target (what we aim for), and that end result is good and clean.
    2. Switch back to starting point, and do:
    3. git diff -R target > diff             # diff to target
      
    4. Edit diff file, to select only those changes we want to include in a first commit. Then do a git-apply diff
    5. vi diff
      git-apply diff                        # Must be in project root dir
      
    6. Test, finalize the last changes before commits, and diff against target if necessary.
    7. # test test test
      git diff -R target > diff             # if necessary
      
    8. Commit, and repeat from step 2.
    9. When done, branch target can be removed
  • Use gitk to get a graphical visualisation of current commit, or some subsets. For instance
  • gitk                                 # View current commit and all ancestors
    gitk master..                        # View changes to current branch (i.e. reachable from HEAD, excluding master)
    
  • Use git stash to save the current state of the working tree (see [25]).
  • git stash                            # Save current work in working tree
    ...                                  # (whatever, including git reset --hard...)
    git stash apply                      # Bring back changes in working tree
    
  • Forgot to add some files in the previous commit? Mistyped the commit message? Use git commit --amend:
  • git commit                           # Oups! forgot one file
    git add somefile                     # ... Add the missing file
    git commit --amend                   # ... and replace the previous commit
    

Word-by-word diffs

Reference:

Add the following alias to your ~/.gitconfig file:

[alias]
    wdiff = diff --color-words
    wshow = show --color-words

Now you can have word-by-word diffs / shows:

   git wdiff fname
   git show

Improving tokenization in diffs

Diffs can even be further improved thanks to the use of the .gitattributes file in your repository. For instance:

*.tex   diff=tex
*.h     diff=cpp
*.c     diff=cpp
*.cpp   diff=cpp

One can also define filters that would clean files before checkin/checkout.

Show Git Branch in Prompt

Links:

The simplest is to add $(__git_ps1) to variable PS1 in file ~/.bashrc. Alternatively one can use \[\e[31m\]$(__git_ps1 " (%s)")\[\e[0m\] to prepad a space when branch name is displayed and use red color.

#PS1 in cygwin
PS1='\[\e]0;\w\a\]\n\[\e[32m\]\u@\h \[\e[33m\]\w\[\e[31m\]$(__git_ps1 " (%s)")\[\e[0m\]\n\$ '

Undoing / Reverting a commit

  • See [28] (stackoverflow.com)

Undoing last commit

The much simpler git commit --amend:

git add ...
git rm ...
git commit --amend          # Current index state will be used for new commit

The usual reset / add / commit sequence:

git commit ...
git reset --soft HEAD^      # Move back to previous commit, keep current changes
# edit edit edit...
git add ...                 # Add modified files
git commit -c ORIG_HEAD     # Commit, reedit old comment as base (-C to reuse w/o editing)

Reverting a past commit

Use git revert ...:

git revert -n HEAD             # Revert last commit, don't commit yet
git revert -n 1234567          # Revert a given commit, don't commit yet
git commit ...

Revert with Cherry Picking changes

The easiest is to reuse the patch method as explained by Linus above in Working the Git Way:

git show -b -R <commit> > diff   # Get a revert patch
vi diff                          # Edit the patch as necessary
git apply diff                   # Apply the patch

Housekeeping

Repack repository:

git repack -adf
git repack -adF                  # Also repack already packed objects - rarely needed

# Complete sequence include garbage collect
rm -rf .git/refs/original/            # remove the old reference backup used for rebase
git reflog expire --expire=now --all  # tell git that all reflog objects expire now
git gc --prune=now                    # first prune everything
git gc --aggressive --prune=now       # ... then again and pack objects together

Squashing Merge Commits

Let's assume we have this situation, and we want to squash the 2 merge commits M1 and M2 (see [29]).

Initial situation Final situation
   X --- Y --------- M1 -------- M2 (my-feature)
  /                 /           /
 /                 /           /
a --- b --- c --- d --- e --- f (stable)
   X --- Y --------------------- M2' (my-feature)
  /                             /
 /                             /
a --- b --- c --- d --- e --- f (stable)

Easy solution:

git checkout my-feature
git reset --soft Y
git rev-parse f > .git/MERGE_HEAD
git commit

Result is commit M2', identical to M2 above.

Finding files

Using git ls-files it is very easy and fast to locate a file in a git repository:

git ls-files                          # List all files in the repository
git ls-files *.c                      # List all files with .d suffix
git ls-files */gcc.mk                 # Find a specific file

Bash Script-Fu

Some tips on how to use Git in scripts.

It is recommended to only use Git's Plumbing and Porcelain commands:

Plumbing
Commands not meant to be used on the command-line, but as building blocks in scripts to build more powerful command
Porcelain
More user-friendly commmands, but that are guaranteed not to change in the future


Get current branch name
The recommended way is to use git symbolic-ref (see [30])
branch="$(git symbolic-ref HEAD 2>/dev/null)" || branch="(no branch)"
branch=${branch##refs/heads/}
Get .git directory location

We can use function __gitdir, but it requires to be exported first. Add to ~/.bashrc or ~/.profile:

export -f __gitdir

Alternatively one can also use [31]:

GITDIR=$(git rev-parse --git-dir)
Get root directory location
GITROOT=$(git rev-parse --show-toplevel)
List all references
#List all references that starts with "refs/remotes/tags"
git-for-each-ref refs/remotes/tags

#Loop on these references
git-for-each-ref refs/remotes/tags | cut -d / -f 4- |
while read REF
do
  echo $REF
done
Detect whether current tree is dirty (from [32])
if ( ! git diff-index --quiet HEAD ); then
    echo Your tree is dirty! Aborting...
    exit 1
fi

Tools

Git-Annex

git-annex allows managing files with git, without checking the file contents into git. While that may seem paradoxical, it is useful when dealing with files larger than git can currently easily handle, whether due to limitations in memory, time, or disk space.

Troubleshoot

Out of memory - fetch

I get the following message under cygwin (WinXP SP3, cygwin 1.7.9-1, git 1.7.5.1-1, perl 5.10.1-5), while doing a git svn fetch:

git svn fetch
# . . .
# . . .
# Out of memory during "large" request for 268439552 bytes, total sbrk() is 311015424 bytes at /usr/lib/perl5/vendor_perl/5.10/Git.pm line 908, <GEN1> line 615.

Solution is to increase Cygwin's maximum memory. Setting /HKLM/Software/Cygwin/heap_chunk_in_mb to 1024 (1GB) fixed the problem. Indeed Cygwin can only allocate 384MB for a process, and process ended up using more than 480MB... (see also [33])

Out of memory - gc / repack

$ git gc --prune --aggressive
Counting objects: 58668, done.
Delta compression using up to 2 threads.
fatal: Out of memory, malloc failed (tried to allocate 65795137 bytes)
error: failed to run repack

Warning! — Seems that git gc --aggressive has it's own settings for repack window. See git-gc.

Settings / configuration we can play on to help in solving that problem:

  • pack.window and pack.depth, which control how objects are stored using delta compression. window basically defines how far repack will look for making deltas, and depth controls the depth when nesting deltas.
  • pack.windowmemory will automatically tune down the values of pack.window and pack.depth when the memory limit is reached. This helps for repositories with both big and small files, allowing to still benefit from a large window with small objects while controling memory usage for big objects.
  • core.bigFileThreshold set the maximum size of objects manipulated for repack. Any object bigger than this threshold will not be repacked and simply stored deflated.
  • pack.threads set the number of threads used by repack (it seems that memory limit is set per thread)
  • Increase the amount of virtual & physical memory
  • Use a 64-bit OS

Settings examples:

  • My settings (got an out of memory - suboptimal pack warning during git gc --prune --aggressive though):
  • pack.windowmemory = 700m                # Default 0 (no limit)
    pack.packsizelimit = 1g                 # Default unlimited
    # No pack.threads settings
    core.bigFileThreshold = 256m            # Default 512m
    
  • Tim's Settings:
  • pack.windowmemory = 256m
    pack.packsizelimit = 1g
    pack.threads = 1
    core.bigFileThreshold = 256m
    
  • Settings on QNAP (ARM-based NAS):
  • sudo git config --system pack.windowmemory 128m
    sudo git config --system pack.packsizelimit 1g
    sudo git config --system pack.threads 1
    sudo git config --system core.bigFileThreshold 256m
    

Some references:

Broken Pipe error code 13 on git svn fetch

This error is most likely due to network disconnection. Solution so far is to fetch repeatedly in a for loop:

for i in 1..50 ; do git svn fetch; done