index : cgit

cgit - CGI for Git
==================

This is an attempt to create a fast web interface for the Git SCM, using a
built-in cache to decrease server I/O pressure.


Installation
------------

Building cgit involves building a proper version of Git. How to do this depends
on how you obtained the cgit sources:

a) If you're working in a cloned cgit repository, you first need to initialize
and update the Git submodule:

    $ git submodule init     # register the Git submodule in .git/config
    $ $EDITOR .git/config    # if you want to specify a different url for git
    $ git submodule update   # clone/fetch and checkout correct git version

b) If you're building from a cgit tarball, you can download a proper git version
like this:

    $ make get-git

When either a) or b) has been performed, you can build and install cgit like
this:

    $ make
    $ sudo make install

This will install `cgit.cgi` and `cgit.css` into `/var/www/htdocs/cgit`. You can
configure this location (and a few other things) by providing a `cgit.conf` file
(see the Makefile for details).

Lua is optional and only powers the lua: filter extensions (authentication,
email and commit-message filters in custom/extensions/). A plain build auto-detects a
Lua through pkg-config, preferring LuaJIT, and falls back to a Lua-less binary
when none is found. Acceptable values are generally "luajit", "lua", "lua5.4",
"lua5.3", "lua5.2" and "lua5.1".

To pin an implementation:

    $ make LUA_PKGCONFIG=lua5.4

To build without Lua, so the binary needs no Lua at runtime:

    $ make NO_LUA=1


Previewing locally
------------------

cgit is a CGI program, so trying it out normally means configuring a web server.
For development there is a small dependency-free preview server at
`tools/serve.py` that runs the built binary and serves the static assets. It
needs only Python 3.

    $ python3 tools/serve.py --config /path/to/cgitrc --port 8080

It is a development aid and is not meant to face the internet.


Dependencies
------------

* zlib
* optional: luajit or lua, most reliably used when pkg-config is available

cgit builds without OpenSSL or libcurl, so no development packages for those are
needed.


Filter extensions
-----------------

The optional Lua filters in `custom/extensions/` need extra Lua modules. Each script's
header lists the exact install commands for its own dependencies.

* The auth filters (`auth-file.lua`, `auth-inline.lua`) need `luaossl` and
  `luaposix`.
* The email filters (`email-gravatar.lua`, `email-libravatar.lua`) need
  `luaossl`.
* The syntax highlighter (`syntax-highlight.lua`) needs `lpeg` and a Scintillua
  lexer set.
* The about-page renderer (`about-render.lua`) needs `lpeg` for markdown and
  man pages. Plain text needs only Lua.

These filters target Lua 5.1 through 5.4 and LuaJIT. `luaossl` has no Lua 5.5
build, so build cgit against 5.1 to 5.4 if you use the auth or email filters.
`link-commits.lua` needs nothing beyond Lua itself.


Web server configuration
------------------------

cgit is a CGI program. Complete, commented configurations for nginx, Apache and
lighttpd live in `custom/servers/`, each explaining how that server routes
requests to the binary and serves the static assets off disk.


URLs and query parameters
-------------------------

cgit serves every page from one CGI program, so a url names the repository, the
page and a path inside that page. There are two forms and they carry the same
information. With virtual-root set, a request is a path, written as
/<repo>/<page>/<path> with anything else as a query string. Without it, the
whole request lives in the query string as ?url=<repo>/<page>/<path>, which is
the form the test suite uses. In the path form the first extra argument opens
the query string with a question mark, and in the url= form it continues the
existing one with an ampersand, which is why links in the two forms are spelled
differently.

The page name is the second element and is one of about, atom, blame, blob,
commit, diff, log, patch, plain, rawdiff, refs, snapshot, stats, summary, tag or
tree. Leaving it out gives the repository summary, and leaving the repository
out as well gives the index of repositories. Everything after the page name is
the path argument, so /demo/tree/src/main.c asks for the tree page at
src/main.c. A repository whose name contains a character that is special in a
url has it percent-encoded, and a plus sign has to be written %2b because a bare
plus decodes to a space.

These parameters are accepted, all of them optional.

    url     repository, page and path in one value, used instead of the
            path form
    h       the branch or ref to read, defaulting to the repository default
    id      pin the page to one commit or object, which every page that
            shows history honours
    id2     the second object for a diff, so id and id2 name the two sides
    ofs     offset into a paged listing, used by log, refs and stats
    path    restrict the page to one path, equivalent to the trailing path
    q       the search term
    qt      what to search, one of grep, author, committer or range
    s       sort key on the index and refs pages
    showmsg show full commit messages in a log listing
    period  the statistics window, one of w, m, q or y
    dt      diff type, selecting unified, side by side or raw
    ss      shorthand for the side by side diff
    all     include every ref rather than one branch, used by atom
    context lines of context in a diff
    ignorews    ignore whitespace when diffing
    follow  follow a single path across renames in a log

A few endpoints are not ordinary pages. The snapshot page takes a filename
rather than a ref, so /demo/snapshot/demo-1.0.tar.gz names both the ref and the
archive format through the suffix, and the formats on offer are set by the
snapshots option. The plain page serves a blob as its own bytes under headers
that stop a browser treating repository content as markup. The atom page is a
feed rather than a page and accepts h, path and all. The clone endpoints under
info and objects implement the dumb HTTP protocol and are only present when http
clone is enabled.

Some worked examples, in the url= form.

    ?url=                              the repository index
    ?url=demo                          summary for the demo repository
    ?url=demo/log&h=next               log of the next branch
    ?url=demo/log&qt=author&q=alice    commits authored by alice
    ?url=demo/tree/src&h=v1.0          the src directory at tag v1.0
    ?url=demo/commit&id=HEAD~3         one commit, pinned
    ?url=demo/diff&id=main&id2=next    diff between two branches
    ?url=demo/plain/README.md          the raw bytes of one file
    ?url=demo/atom&h=main              the commit feed for a branch

Runtime configuration
---------------------

The file `/etc/cgitrc` is read by cgit before handling a request. In addition to
runtime parameters, this file may also contain a list of repositories displayed
by cgit (see `cgitrc.5.txt` for further details). A fully commented starting
point with every option at its default is in `custom/cgitrc`.


Securing an instance
--------------------

A public instance needs a few deliberate choices, all set in cgitrc and
documented in `cgitrc.5.txt`.

* Keep private repositories out of `scan-path`, or set `strict-export` to a
  marker filename so only repositories that contain it are published.

* Gate the whole instance behind a login with `auth-filter`. Two example filters
  ship in `custom/extensions/`, `auth-inline.lua` and `auth-file.lua`.

* Terminate TLS at the web server in front of cgit.

* The example configs in `custom/servers/` set a Content-Security-Policy and
  related headers at the web server, where they also cover the static assets.

* `max-blob-size` bounds how much a single request reads into memory, and
  defaults to 10 MB.

* Leave `enable-cache-list` off, since it exposes the cache path and the URLs
  other visitors requested.

* Build the deployed binary with the hardening flags via
  `tools/release-build.sh`.


The cache
---------

When cgit is invoked it looks for a cache file matching the request and returns
it to the client. If no such cache file exists (or if it has expired), the
content for the request is written into the proper cache file before the file is
returned.

If the cache file has expired but cgit is unable to obtain a lock for it, the
stale cache file is returned to the client. This is done to favour page
throughput over page freshness.

The generated content contains the complete response to the client, including
the HTTP headers.