These tasks are conditional and mutually exclusive due to the "when"
clause. Using import_tasks means that these are imported before the
playbook execution and then skipped during evaluation of the test.
It makes sense in this case to use include_tasks so that the tasks
are only imported during playbook execution if the condition is met.
Ansible 2.4 and 2.5 are moving away from specialized loop functions
and the old syntax will eventually be deprecated and removed. I did
not change the with_fileglob loops because I'm not sure about their
syntax yet.
See: https://docs.ansible.com/ansible/latest/user_guide/playbooks_loops.html
I have zero idea if we have IE6 clients any more, but according to the
H5BP community IE6 actually did support gzip and only represents 0.1%
of Internet traffic in 2015 (!) anyways.
See: https://github.com/h5bp/server-configs-nginx/issues/125
You should only use the "shell" module when you need shell functions
like flow control and redirects. Also, the "command" module is safer
because it is not affected by the user's environment.
Something seems to have happened as of Ansible 2.4.0.0 where this no
longer works. I suspect it is related to the major changes to static
and dynamic imports that landed around this same time.
In practice this achieves the same function, but without the "magic"
ability to use one task for different operating systems.
Something seems to have happened as of Ansible 2.4.0.0 where this no
longer works. I suspect it is related to the major changes to static
and dynamic imports that landed around this same time.
We make sure that this tasks always runs by using the special tag of
the same name.
[ANSIBLE0006] systemctl used in place of systemd module
Also, move the functionality of the changed check to the systemd task,
because it has the ability to simply daemon-reload itself now.
I can't remember right now why I needed to use Debian's MariaDB build
but now I just want to use upstream's latest stable. Debian's version
is 10.1 and upstream has moved on to 10.2.
Ansible 2.4 changes the way includes work. Now you have to use "import"
for playbooks and tasks that are static, and "include" for those that
are dynamic (ie, those that use variables, loops, etc).
See: http://docs.ansible.com/ansible/devel/playbooks_reuse_includes.html
Using www-data was a temporary measure while I was waiting for the
official nginx.org packages to be released for Debian 9 and we had
to use Debian's own nginx package.
Sync with latest packaged version from Debian 9. Effectively this
only updates comments and disabled options, but having less changes
to look at during an apt upgrade is nice and reduces the chance of
adding errors.
There are currently no nginx.org builds for Debian 9, so we need to
use the package from Debian's repository. This package provides a
www-data user and group instead of an nginx one.
We can revert some of this after Debian 9 is released and official
builds come from nginx.org (though it might be useful to keep the
main nginx.conf as a template).
Debian 9 is still in beta and doesn't have nginx.org builds yet, so
we need to use the nginx package in Debian's repositories, and that
required a bit of a different configuration.
After official nginx.org builds are released we can revert this.
I'm surprised this worked all these years actually. Since Ansible
version 1.6 it has been possible to use the version_compare filter
instead of doing math logic on strings.
See: https://docs.ansible.com/ansible/playbooks_tests.html
I realized the other day that due to complex logic in the location
blocks, various WordPress static files like images and stylesheets
didn't get the HTTP Strict Transport Security header set. We need
to include it on each level where we are setting headers, because
nginx overwrites headers if you set them again in a child block.
There was some knowledge floating around that 860 bytes was the
optimal size, I think it was from an Akamai engineer or something,
but the HTML 5 Boilerplate server configs use 256 bytes, and I
actually have HTML content that is less than 860 bytes, so I guess
I could benefit from compressing it. gzip compression is costly
for the compression side, but very quick for the client, so this
is a good thing.
See: https://github.com/h5bp/server-configs-nginx/blob/master/nginx.conf
The variable name is misleading as this really does is install the
certbot client and its dependencies, and we generally want this to
always happen. If a host doesn't want it, they can override it in
their host vars.
Perhaps I should rename this variable to "bootstrap_letsencrypt" or
something so it is more accurate.
The `ConditionFileIsExecutable` goes in the [Unit] section! This
fixes the error:
systemd[1]: [/etc/systemd/system/renew-letsencrypt.service:6] Unknown lvalue 'ConditionFileIsExecutable' in section 'Service'
Only vhosts running WordPress, etc need PHP. Make sure to set the
appopriate variables for each vhost in the host's vars, ie:
nginx_vhosts:
- domain_name: example.com
has_wordpress: True
- domain_name: example.net
needs_php: True
You can set either of them, but not both—needing WordPress implies
needing PHP.
Used to indicate if a vhost needs PHP configuration or not, like
for a static site. Set in the hosts's nginx_vhosts block. Defaults
to "False" if unset.
This reverts commit 201165cff6.
Turns out this actually breaks initial deployments, because the
cache gets updated in the first task, then you add sources for
nginx and mariadb, but it doesn't update the indexes because the
cache is < 3600 seconds old, so you end up getting the distro's
versions of nginx and mariadb.
When you give Ansible the key id it will check if the key exists
before trying to download and add it. I got the long fingerprint
from `sudo apt-key finger`.
When I originally deployed Ubuntu 16.04 there were no public xenial
MariaDB builds yet, so I used wily (15.10). Now there are official
xenial builds so we can use them instead.
I have added cache_valid_time=3600 for the first task in each
tag that could be possibly running apt-related commands. For ex,
the "nginx" tag is also in the "packages" tag, but sometimes you
run the nginx tag by itself (perhaps repeatadely), so you'd want
to limit the update unless the cache was 1 hour old
I never modify file in the git repo, but the WordPress updater does
updates from the web (for example TwentySixteen theme), and this
always causes the task to fail when I switch WordPress versions.
This reverts commit a38d822fad.
The docs definitely recommend twice a day. From a note on certbot's
installation page:
> if you're setting up a cron or systemd job, we recommend running
> it twice per day (it won't do anything until your certificates
> are due for renewal or revoked, but running it regularly would
> give your site a chance of staying online in case a Let's
> Encrypt-initiated revocation happened for some reason). Please
> select a random minute within the hour for your renewal tasks.
See: https://certbot.eff.org/#ubuntuxenial-nginx
I had removed them from Debian 8 and Ubuntu 14.04 configs last year
when the NSA's Suite B crypto guidelines dropped 128-bit algorithms
but those changes didn't make it to my new Ubuntu 16.04 config.
It is probably overkill and paranoid, but this server is mine, so I
can make those decisions (and I only connect from modern clients).
For idempotence we need to run all apt-related tasks, like editing
source files, adding keys, installing packages, etc, when running
the 'packages' tag.
Take an opinionated stance on HTTPS and assume that hosts are using
HTTPS for all vhosts. This can either be via custom TLS cert/key
pairs defined in the host's variables (could even be self-signed
certificates on dev boxes) or via Let's Encrypt.
Hosts can specify use_letsencrypt: 'yes' in their host_vars. For
now this assumes that the certificates already exist (ie, you have
to manually run Let's Encrypt first to register/create the certs).
When you use the apt_repository module it adds a sources.list with
an annoying filename, and also it's just easier to use a template
when we have different distros/versions to support.
The creation of the fastcgi cache dir is part of the nginx role and
should be labled as such. In situations where you only run nginx
tasks with `-t nginx` nginx will fail to start due to the missing
cache dir.
Google's preload check application pointed out that there was an
extra semi colon in the HTTP header:
$ hstspreload checkdomain alaninkenya.org
Warning:
1. Syntax warning: Header includes an empty directive or extra semicolon.
The tool can be downloaded here: https://github.com/chromium/hstspreload
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Google's PageSpeed Insights tool pointed out that the Genericons
in WordPress' Jetpack module could be compressed.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Everything is HTTPS now, whether self-signed or otherwise, so it
doesn't make sense to have a config switch for this.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
It's just deduplication, since it's already obvious that the dict
is for nginx-related vars:
- nginx_domain_name→domain_name
- nginx_domain_aliases→domain_aliases
- nginx_enable_https→enable_https
- nginx_enable_hsts→enable_hsts
Signed-off-by: Alan Orth <alan.orth@gmail.com>
It would be bettwe to set these defaults in the role's defaults, but
we can't because they exist in dicts for each of the host's sites.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
The `enable_https` option in host_vars becomes `nginx_enable_https`
to be more consistent with other nginx options used in host_vars.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
I added another WordPress blog so I need more memory for caching
now. Eventually I wonder if I should deduplicate these somehow...
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Set `use_snakeoil_cert: 'yes'` in host_vars. This is good for dev
hosts where we don't have real domains or real certs. But everything
should have TLS.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Automatically uses the best mirror for your location, see:
http://httpredir.debian.org/demo.html
Should be much better than any hardcoded default for most hosts.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
For now I generated the certs manually, but in the future the play-
book should run the letsencrypt-auto client for us!
Signed-off-by: Alan Orth <alan.orth@gmail.com>
We need to actually check if HSTS was requested before setting the
header in the block handing PHP requests. We check in the main vhost
block, but nginx headers are only inherited if you don't set ANY
headers in child blocks (ie, headers set in parent blocks are cleared
if you set any new ones in the child).
Signed-off-by: Alan Orth <alan.orth@gmail.com>
This is really a per-site setting, so it doesn't make sense to have
a role default. Anyways, HSTS is kinda tricky and potentially dang-
erous, so unless a vhost explicitly sets it to "yes" we shouldn't
enable it.
Note: also switch from using a boolean to using a string; it is st-
ill declarative, but at least now I don't have to guess whether it
is being treated as a bool or not.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
We actually need to use /var/log/munin for munin-node on Debian
too, as that's what is created by the package manager during
installation.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Note: I've only tested this on a Debian container, and you can't
set these sysctls on containers (the host controls them). To make
matters worse, there is no fact to make ansible skip this on hosts
that are running in containers. For now I will just skip it on
hosts that are "virtualization" servers... even though we actually
do have KVM running on Debian on real hardware. *sigh*
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Needed in Ubuntu 15.04 where iptables-persistent is going away. I
have added translations of the current IPv4 and IPv6 iptables rules.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
- Don't run the static files as templates
- Use a separate playbook for related tasks
- Use a template for security.sources.list
Signed-off-by: Alan Orth <alan.orth@gmail.com>
For security and predictability clients should only get a reponse
if they request a hostname we are actually hosting.
If TLS is in use then this will use a self-signed snakeoil cert for
an HTTPS-enabled blank, default vhost.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
New hosts often fail due to not having an apt_mirror, because there
isn't one defined for their group and their host_vars haven't over-
ridden it.
We want new hosts to deploy successfully, so let's just use a default
apt_mirror if there isn't one defined. Rather have a slow mirror than
a failed deployment. And in any case, Linode can download from KENET's
mirror at 10MB/sec. ;)
Signed-off-by: Alan Orth <alan.orth@gmail.com>
I realized there was no need to do a full clone when I was working
in a Vagrant environment in a coffee shop with slow Internet. ;)
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Instead of using dynamic hack to use the package manager for the
current host. We only have Ubuntu here anyways.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
A template is better than ansible's `apt_repository` module because
we can idempotently control the contents of the file based on vari-
ables.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
I was only setting it on the PHP block, which is for all dynamic
requests (ie pages from WordPress), but it should also be the same
for all static files not served from that block.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Include subdomains in the HTTP Strict Transport Security header,
and include the "preload" verb to inform Google we want to be pre-
loaded into the HSTS preload.
See: https://hstspreload.appspot.com/
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Reduce memory allocation from 128 -> 72M because after a few days
of running it's only using 64 or so, so it's really just a waste of
memory.
Also, disable opcache for CLI. What the hell do you need opcaching
in the CLI invocation for? It only persists for one process!
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Adds a default php.ini for php5-fpm from Ubuntu 14.04 which enables
sane settings for PHP 5.5's opcache as well as disables pathinfo.
Closes#9.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
I was attempting to make the config easier to use in test environments
where the key is self-signed, but meh, I rarely do that and I think
this logic doesn't actually work.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
nginx inherits headers from higher-level blocks UNLESS we also set
headers in the current block. In this case the FastCGI cache header
was being set, so we weren't getting the extra-security ones.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
nginx blocks inherit headers set in blocks above them UNLESS the
current level also sets headers[0]. This was causing PHP requests
to not have STS headers because of the FastCGI cache header which
is set in that block.
[0] http://nginx.org/en/docs/http/ngx_http_headers_module.html
Fixes GitHub #7.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
All servers with non-rotating disks (SSDs) should be running noop,
and the rest should be running deadline.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
I think it was originally supposed to be `ansible_os_family` but
we don't have anything other than Ubuntu, so let's just use that.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
nginx is caching HEAD requests, then when users come along and do
a GET request they get an HTTP 200 with no request body. It seems
setting fastcgi_request_methods to GET doesn't stop nginx from caching
HEADs, so for now just add the $request_method to the key.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
It's just too tricky to manage this. Ubuntu / RedHat preseeds and
kickstarts can create the user and add it to groups, but only when
we control the initial boot environment (ie not on Linode, Digital
Ocean, etc), so let's just say we assume this user exists and can
get root with sudo by the some we are running ansible on it.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Bypasses caching for logged in users (right now only for sessions
where the "wordpress_logged_in" cookie is set. Doubles the trans-
actions per second as measured by siege:
$ siege -d1 -t1M -c50 https://mjanja.ch
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Need to make sure the user gets created on a fresh install, like on
Amazon EC2 or OpenStack images where the first user is `ubuntu' and
you can't assume `provisioning' is already created.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Reduces round trip time for clients. Note: I am using a certificate
chain in the `ssl_certificate' directive, so as I understand it, I
don't need to use an explicit trusted intermediate + root CA cert
with the `ssl_trusted_certificate' option. See the nginx docs for
more[0]. Addresses GitHub Issue #5.
Seems to be working, test with:
$ openssl s_client -connect mjanja.ch:443 -servername mjanja.ch -tls1 -tlsextdebug -status
Look for "OCSP Response" with "Cert Status: good".
[0] http://nginx.org/en/docs/http/ngx_http_ssl_module.html#ssl_stapling
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Default is 5 minutes, but it seems like unless you're a high-traff-
ic site, there's no need to expire sessions so quickly. Also, the
istlsfastyet.com configs are using 24 hours, so surely we can.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Use "public" with "max-age" instead of Expires, as "max-age" is always
preferred if it's present. Note: setting "public" doesn't make the
resource "more cacheable", but it is just more explicit.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
My WordPress blogs have a /wordpress subdirectory in the document
root, but I don't serve from the /wordpress URI.
Technically, all we need is the tweaks to the try_files:
- `?args` passes query strings to php5-fpm
- removing 404 from the vhost's try_files so we don't return 404
when the requested file doesn't exist (obviously not all request
URI's in WordPress are actual files on the disk)
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Assumes you have a TLS cert for one domain, but not the others, ie:
http://blah.com \
http://blah.net -> https://blah.iohttp://blah.org /
Otherwise, without https, it creates a vhost with all domain names.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Ansible's mysql module can get this password and connect fine, but
`mysql` on the command line chokes if the password is slightly
complicated and is not quoted.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Relies on the host having a dict with appropriate values defined,
for example:
wordpress_blogs:
- site_name: blah.com
wordpress_version: 3.9.2
wordpress_db_user: db_user
wordpress_db_name: db_name
wordpress_db_pass: V9&XvvKu1hYl
Signed-off-by: Alan Orth <alan.orth@gmail.com>
Without this, all requests to directory URIs throw 403 errors due
to directory listings not being allowed.
Signed-off-by: Alan Orth <alan.orth@gmail.com>
