Neilj/1711faq (#4572)

MSC1711 certificates FAQ

3 files changed, 263 insertions, 0 deletions

diff --git a/UPGRADE.rst b/UPGRADE.rst
index c46f70f699..7bd631f14c 100644
--- a/UPGRADE.rst
+++ b/UPGRADE.rst
@@ -80,6 +80,8 @@ from Let's Encrypt, depending on your server configuration. Of course, if you
 already have a valid certificate for your homeserver's domain, that can be
 placed in Synapse's config directory without the need for ACME.
 
+For more information on configuring TLS certificates see the `FAQ <https://github.com/matrix-org/synapse/blob/master/docs/MSC1711_certificates_FAQ.md>`_
+
 Upgrading to v0.34.0
 ====================
 
diff --git a/changelog.d/4572.misc b/changelog.d/4572.misc
new file mode 100644
index 0000000000..ea5d49b706
--- /dev/null
+++ b/changelog.d/4572.misc
@@ -0,0 +1 @@
+FAQ to help server admins configure TLS certs in 0.99.0
diff --git a/docs/MSC1711_certificates_FAQ.md b/docs/MSC1711_certificates_FAQ.md
new file mode 100644
index 0000000000..0dcef57733
--- /dev/null
+++ b/docs/MSC1711_certificates_FAQ.md
@@ -0,0 +1,260 @@
+# MSC 1711 Certificates FAQ
+
+The goal of Synapse 0.99.0 is to act as a stepping stone to Synapse 1.0.0. It
+supports the r0.1 release of the server to server specification, but is
+compatible with both the legacy Matrix federation behaviour (pre-r0.1) as well
+as post-r0.1 behaviour, in order to allow for a smooth upgrade across the
+federation.
+
+The most important thing to know is that Synapse 1.0.0 will require a valid TLS
+certificate on federation endpoints. Self signed certificates will not be
+sufficient.
+
+Synapse 0.99.0 makes it easy to configure TLS certificates and will
+interoperate with both >= 1.0.0 servers as well as existing servers yet to
+upgrade.
+
+**It is critical that all admins upgrade to 0.99.0 and configure a valid TLS
+certificate.** Admins will have 1 month to do so, after which 1.0.0 will be
+released and those servers without a valid certificate will not longer be able
+to federate with >= 1.0.0 servers.
+
+If you are unable to generate a valid TLS certificate for your server (e.g.
+because you run it on behalf of someone who doesn't want to give you a TLS
+certificate for their domain, or simply because the matrix domain is hosted on
+a different server), then you can now create a /.well-known/matrix/server file
+on the matrix domain in order to delegate Matrix hosting to another domain.
+ Admins who currently use SRV records to delegate a domain **which they do not
+control TLS for** will need to switch to using .well-known/matrix/server - though
+they should retain their SRV record while the federation upgrades over the
+course of the month.  Other SRV records are unaffected.
+
+Full upgrade notes can be found in
+[UPGRADE.rst](https://github.com/matrix-org/synapse/blob/master/UPGRADE.rst).
+What follows is a timeline and some frequently asked questions.
+
+For more details and context on the release of the r0.1 Server/Server API and
+imminent Matrix 1.0 release, you can also see our
+[main talk from FOSDEM 2019](https://matrix.org/blog/2019/02/04/matrix-at-fosdem-2019/).
+
+## Contents
+* Timeline
+* Synapse 0.99.0 has just been released, what do I need to do right now?
+* How do I upgrade?
+* What will happen if I do not set up a valid federation certificate
+  immediately?
+* What will happen if I do nothing at all?
+* When do I need a SRV record or .well-known URI?
+* Can I still use an SRV record?
+* I have created a .well-known URI. Do I still need an SRV record?
+* It used to work just fine, why are you breaking everything?
+* Can I manage my own certificates rather than having Synapse renew
+  certificates itself?
+* Do you still recommend against using a reverse-proxy on the federation port?
+* Do I still need to give my TLS certificates to Synapse if I am using a
+  reverse-proxy?
+* Do I need the same certificate for the client and federation port?
+* How do I tell Synapse to reload my keys/certificates after I replace them?
+
+
+### Timeline
+
+**5th Feb 2019  - Synapse 0.99.0 is released.**
+
+All server admins are encouraged to upgrade.
+
+0.99.0:
+
+-   provides support for ACME to make setting up Let's Encrypt certs easy, as
+    well as .well-known support.
+
+-   does not enforce that a valid CA cert is present on the federation API, but
+    rather makes it easy to set one up.
+
+-   provides support for .well-known
+
+Admins should upgrade and configure a valid CA cert. Homeservers that require a
+.well-known entry (see below), should retain their SRV record and use it
+alongside their .well-known record.
+
+**>= 5th March 2019  - Synapse 1.0.0 is released**
+
+1.0.0 will land no sooner than 1 month after 0.99.0, leaving server admins one
+month after 5th February to upgrade to 0.99.0 and deploy their certificates. In
+accordance with the the [S2S spec](https://matrix.org/docs/spec/server_server/r0.1.0.html)
+1.0.0 will enforce federation checks. This means that any homeserver without a
+valid certificate after this point will no longer be able to federate with
+1.0.0 servers.
+
+### Synapse 0.99.0 has just been released, what do I need to do right now?
+
+Upgrade as soon as you can in preparation for Synapse 1.0.0.
+
+### How do I upgrade?
+
+Follow the upgrade notes here [UPGRADE.rst](https://github.com/matrix-org/synapse/blob/master/UPGRADE.rst)
+
+### What will happen if I do not set up a valid federation certificate immediately?
+
+Nothing initially, but once 1.0.0 is in the wild it will not be possible to
+federate with 1.0.0 servers.
+
+### What will happen if I do nothing at all?
+
+If the admin takes no action at all, and remains on a Synapse < 0.99.0 then the
+homeserver will be unable to federate with those who have implemented
+.well-known. Then, as above, once the month upgrade window has expired the
+homeserver will not be able to federate with any Synapse >= 1.0.0
+
+### When do I need a SRV record or .well-known URI?
+
+If your homeserver listens on the default federation port (8448), and your
+server_name points to the host that your homeserver runs on, you do not need an
+SRV record or .well-known/matrix/server URI.\
+For instance, if you registered example.com and pointed its DNS A record at a
+fresh Upcloud VPS or similar, you could install Synapse 0.99 on that host,
+giving it a server_name of example.com, and it would automatically generate a
+valid TLS certificate for you via Let's Encrypt and no SRV record or
+.well-known URI would be needed.
+
+This is the common case, although you can add an SRV record or
+.well-known/matrix/server URI for completeness if you wish.
+
+**However**, if your server does not listen on port 8448, or if your server_name
+does not point to the host that your homeserver runs on, you will need to let
+other servers know how to find it.
+
+The easiest way to do this is with a .well-known/matrix/server URI on the
+webroot of the domain to advertise your server. For instance, if you ran
+"matrixhosting.com" and you were hosting a Matrix server for example.com, you
+would ask example.com to create a file at:
+
+`https://example.com/.well-known/matrix/server`
+
+with contents:
+
+`{"m.server": "example.matrixhosting.com:8448"}`
+
+...which would tell servers trying to connect to example.com to instead connect
+to example.matrixhosting.com on port 8448. You would then configure Synapse
+with a server_name of "example.com", but generate a TLS certificate for
+example.matrixhosting.com.
+
+As an alternative, you can still use an SRV DNS record for the delegation, but
+this will require you to have a certificate for the matrix domain (example.com
+in this example). See "Can I still use an SRV record?".
+
+### Can I still use an SRV record?
+
+Firstly, if you didn't need an SRV record before (because your server is
+listening on port 8448 of your server_name), you certainly don't need one now:
+the defaults are still the same.
+
+If you previously had an SRV record, you can keep using it provided you are
+able to give Synapse a TLS certificate corresponding to your server name. For
+example, suppose you had the following SRV record, which directs matrix traffic
+for example.com to matrix.example.com:443:
+
+```
+_matrix._tcp.example.com. IN SRV 10 5 443 matrix.example.com
+```
+
+In this case, Synapse must be given a certificate for example.com - or be
+configured to acquire one from Let's Encrypt.
+
+If you are unable to give Synapse a certificate for your server_name, you will
+also need to use a .well-known URI instead. However, see also "I have created a
+.well-known URI. Do I still need an SRV record?".
+
+### I have created a .well-known URI. Do I still need an SRV record?
+
+As of Synapse 0.99, Synapse will first check for the existence of a .well-known
+URL and follow any delegation it suggests. It will only then check for the
+existence of an SRV record.
+
+That means that the SRV record will often be redundant. However, you should
+remember that there may still be older versions of Synapse in the federation
+which do not understand .well-known URIs, so if you removed your SRV record you
+would no longer be able to federate with them.
+
+It is therefore best to leave the SRV record in place for now. Synapse 0.34 and
+earlier will follow the SRV record (and not care about the invalid
+certificate). Synapse 0.99 and later will follow the .well-known URI, with the
+correct certificate chain.
+
+### It used to work just fine, why are you breaking everything?
+
+We have always wanted Matrix servers to be as easy to set up as possible, and
+so back when we started federation in 2014 we didn't want admins to have to go
+through the cumbersome process of buying a valid TLS certificate to run a
+server. This was before Let's Encrypt came along and made getting a free and
+valid TLS certificate straightforward. So instead, we adopted a system based on
+[Perspectives](https://en.wikipedia.org/wiki/Convergence_(SSL)): an approach
+where you check a set of "notary servers" (in practice, homeservers) to vouch
+for the validity of a certificate rather than having it signed by a CA. As long
+as enough different notaries agree on the certificate's validity, then it is
+trusted.
+
+However, in practice this has never worked properly. Most people only use the
+default notary server (matrix.org), leading to inadvertent centralisation which
+we want to eliminate. Meanwhile, we never implemented the full consensus
+algorithm to query the servers participating in a room to determine consensus
+on whether a given certificate is valid. This is fiddly to get right
+(especially in face of sybil attacks), and we found ourselves questioning
+whether it was worth the effort to finish the work and commit to maintaining a
+secure certificate validation system as opposed to focusing on core Matrix
+development.
+
+Meanwhile, Let's Encrypt came along in 2016, and put the final nail in the
+coffin of the Perspectives project (which was already pretty dead). So, the
+Spec Core Team decided that a better approach would be to mandate valid TLS
+certificates for federation alongside the rest of the Web. More details can be
+found in
+[MSC1711](https://github.com/matrix-org/matrix-doc/blob/master/proposals/1711-x509-for-federation.md#background-the-failure-of-the-perspectives-approach).
+
+This results in a breaking change, which is disruptive, but absolutely critical
+for the security model. However, the existence of Let's Encrypt as a trivial
+way to replace the old self-signed certificates with valid CA-signed ones helps
+smooth things over massively, especially as Synapse can now automate Let's
+Encrypt certificate generation if needed.
+
+### Can I manage my own certificates rather than having Synapse renew certificates itself?
+
+Yes, you are welcome to manage your certificates yourself. Synapse will only
+attempt to obtain certificates from Let's Encrypt if you configure it to do
+so.The only requirement is that there is a valid TLS cert present for
+federation end points.
+
+### Do you still recommend against using a reverse-proxy on the federation port?
+
+We no longer actively recommend against using a reverse proxy. Many admins will
+find it easier to direct federation traffic to a reverse-proxy and manage their
+own TLS certificates, and this is a supported configuration.
+
+### Do I still need to give my TLS certificates to Synapse if I am using a reverse proxy?
+
+Practically speaking, this is no longer necessary.
+
+If you are using a reverse-proxy for all of your TLS traffic, then you can set
+`no_tls: True`. In that case, the only reason Synapse needs the certificate is
+to populate a legacy 'tls_fingerprints' field in the federation API. This is
+ignored by Synapse 0.99.0 and later, and the only time pre-0.99 Synapses will
+check it is when attempting to fetch the server keys - and generally this is
+delegated via `matrix.org`, which is on 0.99.0.
+
+However, there is a bug in Synapse 0.99.0
+[4554](<https://github.com/matrix-org/synapse/issues/4554>) which prevents
+Synapse from starting if you do not give it a TLS certificate. To work around
+this, you can give it any TLS certificate at all. This will be fixed soon.
+
+### Do I need the same certificate for the client and federation port?
+
+No. There is nothing stopping you doing so, particularly if you are using a
+reverse-proxy. However, Synapse will use the same certificate on any ports
+where TLS is configured.
+
+### How do I tell Synapse to reload my keys/certificates after I replace them?
+
+Synapse will reload the keys and certificates when it receives a SIGHUP - for
+example kill -HUP $(cat homeserver.pid). Alternatively, simply restart Synapse,
+though this will result in downtime while it restarts.