One of the many ways to provide TLS-encrypted HTTP access (a.k.a. HTTPS) to Fossil is to run it behind a web proxy that supports TLS. This document explains how to use the powerful nginx web server to do that.
Benefits
This scheme is complicated, even with the benefit of this guide and pre-built binary packages. Why should you put up with this complexity? Because it gives many benefits that are difficult or impossible to get with the less complicated options:
Power — nginx is one of the most powerful web servers in the world. The chance that you will run into a web serving wall that you can’t scale with nginx is very low.
To give you some idea of the sort of thing you can readily accomplish with nginx, your author runs a single public web server that provides transparent name-based virtual hosting for four separate domains:
- One is entirely static, not involving any dynamic content or Fossil integration at all.
- Another is served almost entirely by Fossil, with a few select static content exceptions punched past Fossil, which are handled entirely via nginx.
- The other two domains are aliases for one another — e.g.
example.com
andexample.net
— with most of the content being static. This pair of domains has three different Fossil repo proxies attached to various sections of the URI hierarchy.
All of this is done with minimal configuration repetition between the site configurations.
Integration — Because nginx is so popular, it integrates with many different technologies, and many other systems integrate with it in turn. This makes it great middleware, sitting between the outer web world and interior site services like Fossil. It allows Fossil to participate seamlessly as part of a larger web stack.
Availability — nginx is already in most operating system binary package repositories, so you don’t need to go out of your way to get it.
Fossil Remote Access Methods
Fossil provides four major ways to access a repository it’s serving remotely, three of which are straightforward to use with nginx:
HTTP — Fossil has a built-in HTTP server:
fossil server
. While this method is efficient and it’s possible to use nginx to proxy access to another HTTP server, this option is overkill for our purposes. nginx is itself a fully featured HTTP server, so we will choose in this guide not to make nginx reinterpret Fossil’s implementation of HTTP.CGI — This method is simple but inefficient, because it launches a separate Fossil instance on every HTTP hit.
Since Fossil is a relatively small self-contained program, and it’s designed to start up quickly, this method can work well in a surprisingly large number of cases.
Nevertheless, we will avoid this option in this document because we’re already buying into a certain amount of complexity here in order to gain power. There’s no sense in throwing away any of that hard-won performance on CGI overhead.
SCGI — The SCGI protocol provides the simplicity of CGI without its performance problems.
SSH — This method exists primarily to avoid the need for HTTPS in the first place. There is probably a way to get nginx to proxy Fossil to HTTPS via SSH, but it would be pointlessly complicated.
SCGI it is, then.
Installing
The first step is to install the pieces we’ll be working with. This varies on different operating systems, so to avoid overcomplicating this guide, we’re going to assume you’re using Ubuntu Server 18.04 LTS, a common Tier 1 offering for virtual private servers.
SSH into your server, then say:
$ sudo apt install certbot fossil nginx
For other operating systems, simply visit the front Certbot web page and tell it what OS and web stack you’re using. Chances are good that they’ve got a good guide for you already.
Running Fossil in SCGI Mode
You presumably already have a working Fossil configuration on the public server you’re trying to set up and are just following this guide to replace HTTP service with HTTPS.
(You can adjust the advice in this guide to get both HTTP and HTTPS service on the same site, but I strongly recommend that you do not do that: the good excuses remaining for continuing to allow HTTP on public web servers are running thin these days.)
I run my Fossil SCGI server instances with a variant of the fslsrv
shell script currently hosted in the Fossil source
code repository. You’ll want to download that and make a copy of it, so
you can customize it to your particular needs.
This script allows running multiple Fossil SCGI servers, one per
repository, each bound to a different high-numbered localhost
port, so
that only nginx can see and proxy them out to the public. The
“example
” repo is on TCP port localhost:12345, and the “foo
” repo is
on localhost:12346.
As written, the fslsrv
script expects repositories to be stored in the
calling user’s home directory under ~/museum
, because where else do
you keep Fossils?
That home directory also needs to have a directory to hold log files,
~/log/fossil/*.log
. Fossil doesn’t put out much logging, but when it
does, it’s better to have it captured than to need to re-create the
problem after the fact.
The use of --baseurl
in this script lets us have each Fossil
repository mounted in a different location in the URL scheme. Here, for
example, we’re saying that the “example
” repository is hosted under
the /code
URI on its domains, but that the “foo
” repo is hosted at
the top level of its domain. You’ll want to do something like the
former for a Fossil repo that’s just one piece of a larger site, but the
latter for a repo that is basically the whole point of the site.
You might also want another script to automate the update, build, and deployment steps for new Fossil versions:
#!/bin/sh
cd $HOME/src/fossil/trunk
fossil up
make -j11
killall fossil
sudo make install
fslsrv
The killall fossil
step is needed only on OSes that refuse to let you
replace a running binary on disk.
As written, the fslsrv
script assumes a Linux environment. It expects
/bin/bash
to exist, and it depends on non-POSIX tools like pgrep
.
It should not be difficult to port to systems like macOS or the BSDs.
Configuring Let’s Encrypt, the Easy Way
If your web serving needs are simple, Certbot can configure nginx for you and keep its certificates up to date. You can follow the Certbot documentation for nginx on Ubuntu 18.04 LTS guide as-is, though we’d recommend one small change: to use the version of Certbot in the Ubuntu package repository rather than the first-party Certbot package that the guide recommends.
The primary local configuration you need is to tell nginx how to proxy
certain URLs down to the Fossil instance you started above with the
fslsrv
script:
location / {
include scgi_params;
scgi_pass 127.0.0.1:12345;
scgi_param HTTPS "on";
scgi_param SCRIPT_NAME "";
}
The TCP port number in that snippet is the key: it has to match the port
number generated by fslsrv
from the base port number passed to the
start_one
function.
Configuring Let’s Encrypt, the Hard Way
If you’re finding that you can’t get certificates to be issued or
renewed using the Easy Way instructions, the problem is usually that
your nginx configuration is too complicated for Certbot’s --nginx
plugin to understand. It attempts to rewrite your nginx configuration
files on the fly to achieve the renewal, and if it doesn’t put its
directives in the right locations, the domain verification can fail.
Let’s Encrypt uses the Automated Certificate Management
Environment protocol (ACME) to determine whether a given client
actually has control over the domain(s) for which it wants a certificate
minted. Let’s Encrypt will not blithely let you mint certificates for
google.com
and paypal.com
just because you ask for it!
Your author’s configuration, glossed above, is complicated enough that the current version of Certbot (0.28 at the time of this writing) can’t cope with it. That’s the primary motivation for me to write this guide: I’m addressing the “me” years hence who needs to upgrade to Ubuntu 20.04 or 22.04 LTS and has forgotten all of this stuff. 😉
Step 1: Shifting into Manual
The first thing to do is to turn off all of the Certbot automation, because it’ll only get in our way. First, disable the Certbot package’s automatic background updater:
$ sudo systemctl disable certbot.timer
Next, edit /etc/letsencrypt/renewal/example.com.conf
to disable the
nginx plugins. You’re looking for two lines setting the “install” and
“auth” plugins to “nginx”. You can comment them out or remove them
entirely.
Step 2: Configuring nginx
On Ubuntu systems, at least, the primary user-level configuration file
is /etc/nginx/sites-enabled/default
. For a configuration like I
described at the top of this article, I recommend that this file contain
only a list of include statements, one for each site that server hosts:
include local/example
include local/foo
Those files then each define one domain’s configuration. Here,
/etc/nginx/local/example
contains the configuration for
*.example.com
and *.example.net
; and local/foo
contains the
configuration for *.foo.net
.
Here’s an example configuration:
server {
server_name .foo.net;
include local/tls-common;
charset utf-8;
access_log /var/log/nginx/foo.net-https-access.log;
error_log /var/log/nginx/foo.net-https-error.log;
# Bypass Fossil for the static Doxygen docs
location /doc/html {
root /var/www/foo.net;
location ~* \.(html|ico|css|js|gif|jpg|png)$ {
expires 7d;
add_header Vary Accept-Encoding;
access_log off;
}
}
# Redirect everything else to the Fossil instance
location / {
include scgi_params;
scgi_pass 127.0.0.1:12345;
scgi_param HTTPS "on";
scgi_param SCRIPT_NAME "";
}
}
server {
server_name .foo.net;
root /var/www/foo.net;
include local/http-certbot-only;
access_log /var/log/nginx/foo.net-http-access.log;
error_log /var/log/nginx/foo.net-http-error.log;
}
Notice that we need two server { }
blocks: one for HTTPS service, and
one for HTTP-only service:
HTTP over TLS (HTTPS) Service
The first server { }
block includes this file, local/tls-common
:
listen 443 ssl;
ssl_certificate /etc/letsencrypt/live/example.com/fullchain.pem;
ssl_certificate_key /etc/letsencrypt/live/example.com/privkey.pem;
ssl_dhparam /etc/letsencrypt/ssl-dhparams.pem;
ssl_stapling on;
ssl_stapling_verify on;
ssl_protocols TLSv1 TLSv1.1 TLSv1.2;
ssl_ciphers "ECDHE-ECDSA-AES128-GCM-SHA256:ECDHE-ECDSA-AES256-GCM-SHA384:ECDHE-ECDSA-AES128-SHA:ECDHE-ECDSA-AES256-SHA:ECDHE-ECDSA-AES128-SHA256:ECDHE-ECDSA-AES256-SHA384:ECDHE-RSA-AES128-GCM-SHA256:ECDHE-RSA-AES256-GCM-SHA384:ECDHE-RSA-AES128-SHA:ECDHE-RSA-AES256-SHA:ECDHE-RSA-AES128-SHA256:ECDHE-RSA-AES256-SHA384:DHE-RSA-AES128-GCM-SHA256:DHE-RSA-AES256-GCM-SHA384:DHE-RSA-AES128-SHA:DHE-RSA-AES256-SHA:DHE-RSA-AES128-SHA256:DHE-RSA-AES256-SHA256”;
ssl_session_cache shared:le_nginx_SSL:1m;
ssl_prefer_server_ciphers on;
ssl_session_timeout 1440m;
These are the common TLS configuration parameters used by all domains hosted by this server.
The first line tells nginx to accept TLS-encrypted HTTP connections on
the standard HTTPS port. It is the same as listen 443; ssl on;
in
older versions of nginx.
Since all of those domains share a single TLS certificate, we reference
the same example.com/*.pem
files written out by Certbot with the
ssl_certificate*
lines.
The ssl_dhparam
directive isn’t strictly required, but without it, the
server becomes vulnerable to the Logjam attack because some of
the cryptography steps are precomputed, making the attacker’s job much
easier. The parameter file this directive references should be
generated automatically by the Let’s Encrypt package upon installation,
making those parameters unique to your server and thus unguessable. If
the file doesn’t exist on your system, you can create it manually, so:
$ sudo openssl dhparam -out /etc/letsencrypt/dhparams.pem 2048
Beware, this can take a long time. On a shared Linux host I tried it on running OpenSSL 1.1.0g, it took about 21 seconds, but on a fast, idle iMac running LibreSSL 2.6.5, it took 8 minutes and 4 seconds!
The next section is also optional. It enables OCSP stapling, a protocol that improves the speed and security of the TLS connection negotiation.
The next section containing the ssl_protocols
and ssl_ciphers
lines
restricts the TLS implementation to only those protocols and ciphers
that are currently believed to be safe and secure. This section is the
one most prone to bit-rot: as new attacks on TLS and its associated
technologies are discovered, this configuration is likely to need to
change. Even if we fully succeed in keeping this document
up-to-date, the nature of this guide is to recommend static
configurations for your server. You will have to keep an eye on this
sort of thing and evolve your local configuration as the world changes
around it.
Running a TLS certificate checker against your site occasionally is a
good idea. The most thorough service I’m aware of is the Qualys SSL
Labs Test, which gives the site I’m basing this guide on an “A”
rating at the time of this writing. The long ssl_ciphers
line above is
based on their advice: the default nginx configuration tells
OpenSSL to use whatever ciphersuites it considers “high security,” but
some of those have come to be considered “weak” in the time between that
judgement and the time of this writing. By explicitly giving the list of
ciphersuites we want OpenSSL to use within nginx, we can remove those
that become considered weak in the future.
There are a few things you can do to get an even better grade, such as to enable HSTS, which prevents a particular variety of man in the middle attack where our HTTP-to-HTTPS permanent redirect is intercepted, allowing the attacker to prevent the automatic upgrade of the connection to a secure TLS-encrypted one. I didn’t enable that in the configuration above, because it is something a site administrator should enable only after the configuration is tested and stable, and then only after due consideration. There are ways to lock your users out of your site by jumping to HSTS hastily. When you’re ready, there are guides you can follow elsewhere online.
HTTP-Only Service
While we’d prefer not to offer HTTP service at all, we need to do so for two reasons:
The temporary reason is that until we get Let’s Encrypt certificates minted and configured properly, we can’t use HTTPS yet at all.
The ongoing reason is that the Certbot ACME HTTP-01 challenge used by the Let’s Encrypt service only runs over HTTP. This is not only because it has to work before HTTPS is first configured, but also because it might need to work after a certificate is accidentally allowed to lapse, to get that server back into a state where it can speak HTTPS safely again.
So, from the second service { }
block, we include this file to set up
the minimal HTTP service we reqiure, local/http-certbot-only
:
listen 80;
listen [::]:80;
# This is expressed as a rewrite rule instead of an "if" because
# http://wiki.nginx.org/IfIsEvil
#rewrite ^(/.well-known/acme-challenge/.*) $1 break;
# Force everything else to HTTPS with a permanent redirect.
#return 301 https://$host$request_uri;
As written above, this configuration does nothing other than to tell nginx that it’s allowed to serve content via HTTP on port 80 as well.
We’ll uncomment the rewrite
and return
directives below, when we’re
ready to begin testing.
Why the Repetition?
These server { }
blocks contain several directives that have to be
either completely repeated or copied with only trivial changes when
you’re hosting multiple domains from a single server.
You might then wonder, why haven’t I factored some of those directives
into the included files local/tls-common
and
local/http-certbot-only
? Why can’t the HTTP-only server { }
block
above be just two lines? That is, why can I not say:
server_name .foo.net;
include local/http-certbot-only;
Then in local/http-certbot-only
say:
root /var/www/$host;
access_log /var/log/nginx/$host-http-access.log;
error_log /var/log/nginx/$host-http-error.log;
Sadly, nginx doesn’t allow variable subtitution into these particular
directives. As I understand it, allowing that would make nginx slower,
so we must largely repeat these directives in each HTTP server { }
block.
These configurations are, as shown, as small as I know how to get them. If you know of a way to reduce some of this repitition, I solicit your advice.
Step 3: Dry Run
We want to first request a dry run, because Let’s Encrypt puts some rather low limits on how often you’re allowed to request an actual certificate. You want to be sure everything’s working before you do that. You’ll run a command something like this:
$ sudo certbot certonly --webroot --dry-run \
--webroot-path /var/www/example.com \
-d example.com -d www.example.com \
-d example.net -d www.example.net \
--webroot-path /var/www/foo.net \
-d foo.net -d www.foo.net
There are two key options here.
First, we’re telling Certbot to use its --webroot
plugin instead of
the automated --nginx
plugin. With this plugin, Certbot writes the
ACME HTTP-01 challenge files to the static web document root
directory behind each domain. For this example, we’ve got two web
roots, one of which holds documents for two different second-level
domains (example.com
and example.net
) with www
at the third level
being optional. This is a common sort of configuration these days, but
you needn’t feel that you must slavishly imitate it; the other web root
is for an entirely different domain, also with www
being optional.
Since all of these domains are served by a single nginx instance, we
need to give all of this in a single command, because we want to mint a
single certificate that authenticates all of these domains.
The second key option is --dry-run
, which tells Certbot not to do
anything permanent. We’re just seeing if everything works as expected,
at this point.
Troubleshooting the Dry Run
If that didn’t work, try creating a manual test:
$ mkdir -p /var/www/example.com/.well-known/acme-challenge
$ echo hi > /var/www/example.com/.well-known/acme-challenge/test
Then try to pull that file over HTTP — not HTTPS! — as
http://example.com/.well-known/acme-challenge/test
. I’ve found that
using Firefox or Safari is better for this sort of thing than Chrome,
because Chrome is more aggressive about automatically forwarding URLs to
HTTPS even if you requested “http
”.
In extremis, you can do the test manually:
$ telnet foo.net 80
GET /.well-known/acme-challenge/test HTTP/1.1
Host: example.com
HTTP/1.1 200 OK
Server: nginx/1.14.0 (Ubuntu)
Date: Sat, 19 Jan 2019 19:43:58 GMT
Content-Type: application/octet-stream
Content-Length: 3
Last-Modified: Sat, 19 Jan 2019 18:21:54 GMT
Connection: keep-alive
ETag: "5c436ac2-4"
Accept-Ranges: bytes
hi
You type the first two lines at the remote system, plus the doubled “Enter” to create the blank line, and you get something back that hopefully looks like the rest of the text above.
The key bits you’re looking for here are the “hi” line at the end — the document content you created above — and the “200 OK” response code. If you get a 404 or other error response, you need to look into your web server logs to find out what’s going wrong.
Note that it’s important to do this test with HTTP/1.1 when debugging a
name-based virtual hosting configuration like this. Unless you test only
with the primary domain name alias for the server, this test will fail.
Using the example configuration above, you can only use the
easier-to-type HTTP/1.0 protocol to test the foo.net
alias.
If you’re still running into trouble, the log file written by Certbot can be helpful. It tells you where it’s writing it early in each run.
Step 4: Getting Your First Certificate
Once the dry run is working, you can drop the --dry-run
option and
re-run the long command above. (The one with all the --webroot*
flags.) This should now succeed, and it will save all of those flag
values to your Let’s Encrypt configuration file, so you don’t need to
keep giving them.
Step 5: Test It
Edit the local/http-certbot-only
file and uncomment the redirect
and
return
directives, then restart your nginx server and make sure it now
forces everything to HTTPS like it should:
$ sudo systemctl restart nginx
Test ideas:
Visit both Fossil and non-Fossil URLs
Log into the repo, log out, and log back in
Clone via
http
: ensure that it redirects tohttps
, and that subsequentfossil sync
commands go directly tohttps
due to the 301 permanent redirect.
This forced redirect is why we don’t need the Fossil Admin → Access "Redirect to HTTPS on the Login page" setting to be enabled. Not only is it unnecessary with this HTTPS redirect at the front-end proxy level, it would actually cause an infinite redirect loop if enabled.
Step 6: Re-Sync Your Repositories
Now that the repositories hosted by this server are available via HTTPS, you need to tell Fossil about it:
$ cd ~/path/to/checkout
$ fossil sync https://example.com/code
Once that’s done per repository file, all checkouts of that repo will from that point on use the HTTPS URI to sync.
You might wonder if that’s necessary, since we have the automatic HTTP-to-HTTPS redirect on this site now. If you clone or sync one of these nginx-hosted Fossil repositories over an untrustworthy network that allows MITM attacks, that redirect won’t protect you from a sufficiently capable and motivated attacker unless you’ve also gone ahead and enabled HSTS. You can put off the need to enable HSTS by explicitly using HTTPS URIs.
Step 7: Renewing Automatically
Now that the configuration is solid, you can renew the LE cert with the
certbot
command from above without the --dry-run
flag plus a restart
of nginx:
sudo certbot certonly --webroot \
--webroot-path /var/www/example.com \
-d example.com -d www.example.com \
-d example.net -d www.example.net \
--webroot-path /var/www/foo.net \
-d foo.net -d www.foo.net
sudo systemctl restart nginx
I put those commands in a script in the PATH
, then arrange to call that
periodically. Let’s Encrypt doesn’t let you renew the certificate very
often unless forced, and when forced there’s a maximum renewal counter.
Nevertheless, some people recommend running this daily and just letting
it fail until the server lets you renew. Others arrange to run it no
more often than it’s known to work without complaint. Suit yourself.
TLS and web proxying are a constantly evolving technology. This article replaces my earlier effort, which had whole sections that were basically obsolete within about a year of posting it. Two years on, and I was encouraging readers to ignore about half of that HOWTO. I am now writing this document about 3 years later because Let’s Encrypt deprecated key technology that HOWTO depended on, to the point that following that old HOWTO is more likely to confuse than enlighten.
There is no particularly good reason to expect that this sort of thing
will not continue to happen, so this effort is expected to be a living
document. If you do not have commit access on the fossil-scm.org
repository to update this document as the world changes around it, you
can discuss this document on the forum. This document’s author
keeps an eye on the forum and expects to keep this document updated with
ideas that appear in that thread.