# nginx configuration for cgit.
#
# This is a complete nginx.conf rather than a snippet for conf.d or
# sites-enabled, so nothing here is included from elsewhere. Install it and
# reload, or point nginx straight at it to try it out.
#     nginx -t -c /path/to/nginx.conf     # check the syntax
#     nginx -c /path/to/nginx.conf        # run it
#
# Paths assumed below, edit them to match your install.
#   cgit CGI binary   /usr/lib/cgit/cgit.cgi
#   static assets     /usr/share/cgit  (cgit.css cgit.js cgit.png favicon.ico robots.txt)
#   cgit config       /etc/cgitrc
#   public URL        https://git.example.org/  (cgit at the domain root)
#
# cgit is one CGI executable. It learns the repository and the page from
# PATH_INFO and reads page options such as h= and id= from QUERY_STRING, so
# nginx must pass PATH_INFO through to the binary. cgit builds its own link
# base from SCRIPT_NAME. The five static assets are served straight off disk
# and must never be routed through cgit. Passing PATH_INFO through is the
# single most important part of the config below.
#
# nginx cannot run CGI programs itself, so a small bridge called fcgiwrap runs
# the cgit.cgi binary and speaks FastCGI to nginx. Nothing below works until
# that socket exists. On Debian and Ubuntu the packaged systemd socket
# provides /run/fcgiwrap.socket, so enabling it is enough.
#     apt install fcgiwrap
#     systemctl enable --now fcgiwrap.socket
# The socket must be readable by nginx's user. The packaged unit runs fcgiwrap
# as www-data, which nginx also uses on those systems. Without systemd you can
# run
#     spawn-fcgi -s /run/fcgiwrap.socket -M 660 -- /usr/sbin/fcgiwrap
# or run fcgiwrap over TCP and point fastcgi_pass at 127.0.0.1:9000.


# The user nginx drops to after binding the ports. It is www-data on Debian
# and Ubuntu, nginx on RHEL and Fedora, and http on Arch and Alpine. It has to
# match whatever owns the fcgiwrap socket.
user                 www-data;
worker_processes     auto;
pid                  /run/nginx.pid;

# Startup and worker errors. Per-site request logs are set in the vhost.
error_log            /var/log/nginx/error.log warn;

events {
    worker_connections 1024;
}


http {
    # nginx's compiled-in type table knows only text/html, and the usual
    # "include mime.types" would pull in a second file. Exactly five static
    # files are served off disk, so their types are declared here instead and
    # this config keeps standing on its own. Everything else nginx returns
    # comes from cgit, which sets its own Content-Type.
    types {
        text/css                  css;
        text/javascript           js;
        image/png                 png;
        image/vnd.microsoft.icon  ico;
        text/plain                txt;
    }
    default_type application/octet-stream;

    sendfile           on;
    tcp_nopush         on;
    keepalive_timeout  65;

    # Drop the version number from the Server header and from error pages.
    server_tokens      off;

    # cgit pages are large and highly compressible, so this is the cheapest
    # speedup available. text/html is always compressed and cannot be listed.
    # Snapshot tarballs are deliberately absent, since they arrive compressed
    # already and running them through gzip again only burns CPU.
    gzip               on;
    gzip_vary          on;
    gzip_proxied       any;
    gzip_min_length    1024;
    gzip_types         text/css text/javascript text/plain application/atom+xml;


    # Requests carrying a Host header this config does not serve, a raw IP or
    # an invented name from a scanning bot, land in these two blocks and are
    # dropped without a response. cgit keys its page cache on the Host header
    # so clone URLs stay honest, which means every invented hostname reaching
    # it would mint a cache entry of its own and evict a real page to make
    # room. Refusing strangers here keeps the cache to the names the site
    # actually answers to. 444 is nginx shorthand for closing the connection
    # without replying.
    server {
        listen      80 default_server;
        listen      [::]:80 default_server;
        server_name _;
        return      444;
    }

    server {
        listen      443 ssl default_server;
        listen      [::]:443 ssl default_server;
        server_name _;
        # Refuse the TLS handshake itself when the SNI name is unknown, which
        # also spares this block from needing a certificate. Requires nginx
        # 1.19.4 or newer. On older builds point ssl_certificate at any cert,
        # a self-signed one included, and rely on the return below.
        ssl_reject_handshake on;
        # A client can still handshake against a real name and then send some
        # other Host header. Those requests route here after the handshake,
        # past the rejection above, so close them too.
        return      444;
    }


    # Bounce plain HTTP up to HTTPS. Delete this whole server block if you
    # serve plain HTTP only.
    server {
        listen      80;
        listen      [::]:80;
        server_name git.example.org;

        # ACME http-01 challenge files, if you use certbot in webroot mode.
        location ^~ /.well-known/acme-challenge/ {
            root /var/www/html;
        }

        # Everything else moves to HTTPS.
        location / {
            return 301 https://$host$request_uri;
        }
    }


    # The site itself, written for TLS on 443. For a quick plain-HTTP test,
    # change the two listen lines to port 80, delete the redirect block above,
    # and delete the http2, ssl and Strict-Transport-Security lines below.
    # Everything else stays as it is.
    server {
        listen      443 ssl;
        listen      [::]:443 ssl;
        # nginx 1.25.1 and newer. On older builds delete this and write the
        # listen lines as "listen 443 ssl http2;" instead.
        http2       on;
        server_name git.example.org;

        ssl_certificate     /etc/letsencrypt/live/git.example.org/fullchain.pem;
        ssl_certificate_key /etc/letsencrypt/live/git.example.org/privkey.pem;
        ssl_protocols       TLSv1.2 TLSv1.3;
        ssl_ciphers         HIGH:!aNULL:!MD5;
        # Let the client pick, which is the modern advice once the ancient
        # protocol versions are already excluded above.
        ssl_prefer_server_ciphers off;
        # Resumption, so a browser paging through a repository is not made to
        # redo a full handshake on every connection.
        ssl_session_cache   shared:SSL:10m;
        ssl_session_timeout 1d;
        ssl_session_tickets off;

        # Site-wide security headers sit here because they must also cover the
        # static assets nginx serves directly. cgit itself sends only the
        # headers the proxy cannot supply. Those are Status, Content-Type,
        # Content-Length and Content-Disposition on downloads, a no-store
        # Cache-Control on unauthenticated responses, the auth filter's
        # Set-Cookie, and on raw repository bytes a nosniff of its own next to
        # the stricter policy "default-src 'none'". Everything else, this
        # policy included, is the proxy's job.
        #
        # add_header appends and never replaces what cgit sent, so a raw page
        # carries both policies and the browser enforces the stricter one,
        # while the doubled nosniff line is harmless. Keep it that way. Any
        # construct that rewrites response headers here could strip the
        # protection cgit puts on raw repository content.
        #
        # cgit loads only its own /cgit.js and uses inline style on the
        # diffstat bars, so script-src stays self while style-src allows
        # inline. form-action self covers the login form, the only form cgit
        # renders. always applies them to error responses too. If you enable
        # the gravatar or libravatar avatar filter, add its host to img-src,
        # for example https://www.gravatar.com or https://seccdn.libravatar.org.
        add_header Content-Security-Policy "default-src 'self'; script-src 'self'; style-src 'self' 'unsafe-inline'; img-src 'self' data:; object-src 'none'; base-uri 'none'; frame-ancestors 'self'; form-action 'self'" always;
        add_header X-Content-Type-Options "nosniff" always;
        add_header Referrer-Policy "no-referrer" always;

        # The browser features a git viewer never asks for, camera and
        # location among them, are refused outright for everything served
        # here, repository files included.
        add_header Permissions-Policy "accelerometer=(), autoplay=(), camera=(), display-capture=(), geolocation=(), gyroscope=(), magnetometer=(), microphone=(), midi=(), payment=(), picture-in-picture=(), usb=()" always;

        # Cross-origin isolation. The opener policy cuts any window.opener
        # link between cgit and pages that open it, and the resource policy
        # stops other origins from embedding what cgit serves, a hotlinked
        # raw file for example. git clients are not browsers and ignore both,
        # so clone and snapshot downloads keep working. The stricter
        # Cross-Origin-Embedder-Policy is deliberately absent because it
        # would break the avatar filters mentioned above.
        add_header Cross-Origin-Opener-Policy "same-origin" always;
        add_header Cross-Origin-Resource-Policy "same-origin" always;

        # Two years of forced HTTPS. This config already redirects every
        # plain request to TLS, so browsers may as well stop asking. Delete
        # this line if you convert the vhost to plain HTTP, and know it is
        # hard to undo once browsers have seen it.
        add_header Strict-Transport-Security "max-age=63072000; includeSubDomains" always;

        # The document root is the directory that holds the static assets. cgit
        # emits absolute links to /cgit.css and /cgit.png by default, so those
        # files must resolve at the root of the URL space. Pointing root at the
        # asset directory makes /cgit.css map to /usr/share/cgit/cgit.css.
        root /usr/share/cgit;

        # The only request body cgit ever reads is the auth-filter login form,
        # and it stops after 4096 bytes. Nothing else here accepts an upload,
        # so keep the cap far below the nginx default of 1m.
        client_max_body_size 64k;

        access_log /var/log/nginx/cgit.access.log combined;
        error_log  /var/log/nginx/cgit.error.log;

        # Match the assets by their exact root-level names, never by bare
        # extension. cgit routes on PATH_INFO and a repository can hold files
        # ending in .css or .png, so /myrepo/tree/style.css and /myrepo/plain/
        # logo.png are real cgit URLs. A broad extension match would capture
        # those, look for them on disk, and return 404 before cgit could render
        # them. Anchoring the regex at the start of the path matches /cgit.css
        # but not /myrepo/tree/cgit.css, so it can never shadow a repository
        # file. An nginx regex location is matched before the prefix location
        # below, so these assets win for their exact URLs and cgit wins for the
        # rest.
        location ~ ^/(cgit\.css|cgit\.js|cgit\.png|favicon\.ico|robots\.txt)$ {
            expires    30d;
            access_log off;
            try_files  $uri =404;
        }

        # Everything that is not a static asset above is a cgit URL, the repo
        # index, a repository, a page within a repository, a snapshot, a feed.
        location / {
            # The CGI environment. This is normally "include fastcgi_params",
            # which would be a second file, so the list is written out here.
            # Of all of it cgit itself reads only CGIT_CONFIG, PATH_INFO,
            # QUERY_STRING, SCRIPT_NAME, REQUEST_METHOD, CONTENT_LENGTH,
            # HTTP_HOST, HTTPS, SERVER_NAME, SERVER_PORT, HTTP_COOKIE and
            # HTTP_REFERER. The cookie and referer arrive on their own, since
            # nginx forwards request headers as HTTP_* without being asked.

            # The program fcgiwrap runs. It must be the cgit binary itself, not
            # $document_root$fastcgi_script_name, which would try to run a repo
            # path and is the usual cause of a failed request.
            fastcgi_param SCRIPT_FILENAME /usr/lib/cgit/cgit.cgi;

            # cgit builds its link base, the virtual root, from SCRIPT_NAME.
            # The stock parameters set SCRIPT_NAME to the whole request path,
            # which would make cgit prepend that path to every link. Served at
            # the domain root the script has no prefix, so force SCRIPT_NAME
            # empty and cgit uses / as its base. A sub-path install sets it
            # instead, see the end of this file.
            fastcgi_param SCRIPT_NAME  "";

            # How cgit learns the repository and page. At the domain root the
            # whole request path is the PATH_INFO.
            fastcgi_param PATH_INFO    $uri;

            # Page options such as h=branch, id=sha and the snapshot format.
            fastcgi_param QUERY_STRING $query_string;

            # Which config cgit reads. It checks CGIT_CONFIG and falls back to
            # the compiled-in /etc/cgitrc. Setting it makes the location
            # explicit and lets you move cgitrc without recompiling.
            fastcgi_param CGIT_CONFIG  /etc/cgitrc;

            # The browser's Host header, so cgit builds clone URLs against the
            # name the visitor used rather than server_name.
            fastcgi_param HTTP_HOST    $http_host;

            # The real scheme, so cgit builds correct https clone URLs.
            fastcgi_param HTTPS        $https if_not_empty;

            # The routine remainder, needed by fcgiwrap and by the login form.
            fastcgi_param REQUEST_METHOD    $request_method;
            fastcgi_param CONTENT_TYPE      $content_type;
            fastcgi_param CONTENT_LENGTH    $content_length;
            fastcgi_param REQUEST_URI       $request_uri;
            fastcgi_param DOCUMENT_URI      $document_uri;
            fastcgi_param DOCUMENT_ROOT     $document_root;
            fastcgi_param SERVER_PROTOCOL   $server_protocol;
            fastcgi_param REQUEST_SCHEME    $scheme;
            fastcgi_param GATEWAY_INTERFACE CGI/1.1;
            fastcgi_param SERVER_SOFTWARE   nginx/$nginx_version;
            fastcgi_param REMOTE_ADDR       $remote_addr;
            fastcgi_param REMOTE_PORT       $remote_port;
            fastcgi_param SERVER_ADDR       $server_addr;
            fastcgi_param SERVER_PORT       $server_port;
            fastcgi_param SERVER_NAME       $server_name;

            # Hand off to the fcgiwrap socket. A TCP fcgiwrap would use for
            # example 127.0.0.1:9000 here.
            fastcgi_pass unix:/run/fcgiwrap.socket;

            # Large outputs such as snapshot tarballs and blame on big files
            # can take a while, so give cgit room and stream rather than buffer.
            fastcgi_read_timeout 300s;
            fastcgi_buffering off;

            # cgit sends no caching headers of its own, so browsers refetch
            # dynamic pages and cgit's internal cache keeps that cheap. Do not
            # add a blanket expires or Cache-Control in this location. It
            # would fight the no-store cgit puts on the login page and could
            # let one visitor's page be served to another from a shared cache.
        }
    }
}


# To serve cgit at https://git.example.org/cgit/ instead of the root, split
# the URL so SCRIPT_NAME is the prefix and PATH_INFO is the rest. Keep every
# other fastcgi_param from the location above.
#
#     location /cgit/ {
#         fastcgi_split_path_info ^(/cgit)(/.*)$;
#         fastcgi_param SCRIPT_FILENAME /usr/lib/cgit/cgit.cgi;
#         fastcgi_param SCRIPT_NAME     $fastcgi_script_name;
#         fastcgi_param PATH_INFO       $fastcgi_path_info;
#         ...
#         fastcgi_pass  unix:/run/fcgiwrap.socket;
#     }
#
# Serve the assets from the sub-path too, again anchored to the exact names.
#
#     location ~ ^/cgit/(cgit\.css|cgit\.js|cgit\.png|favicon\.ico|robots\.txt)$ {
#         alias /usr/share/cgit/$1;
#         expires 30d;
#         access_log off;
#     }
#
# cgit's default css=/cgit.css and logo=/cgit.png point at the domain root, so
# under a sub-path also set css=/cgit/cgit.css and logo=/cgit/cgit.png in
# cgitrc. cgit derives the /cgit prefix from SCRIPT_NAME. If links come out
# wrong, pin it in cgitrc with virtual-root=/cgit/.