2011 October

October 2011

Avanzada 7 have published the video from my talk at Voip2Day in Madrid in the beginning of October. Enjoy!

In a SIP network, you often have multiple servers communicating with each other. As soon as you add TCP and TLS to the mix, you will want to reuse connections. Why? Setting up A TLS connection involves a lot of messages going back and forth in the process up validating certificates and coming up with keying material for the encrypted session. Now if you have a re-invite that wants to put a call on hold, you don’t want to loose a lot of packet-roundtrip-times while this happens. A better solution is to keep connections open where possible and allow communication both ways.

RFC 3261 states that if you open a connection with a connection-oriented protocol, like TCP or STCP, the connection should stay open to cover the whole transaction. This means that if the other end sends a message in the dialog, a connection needs to be opened in the other direction. This is of course a problem with NAT between a device and a server, something that the SIP Outbound standard handles. Between servers, like B2bua’s and proxys, the problem still exists. This is managed by the Connection Reuse RFC, RFC 5923.

Mutual TLS authentication opens up for two-way communication

RFC 5923 – SIP Connection reuse – explains how this can work. One requirement is that the TLS connection has mutual connection, which means that the server ask the client for a certificate. The client indicates in the request that it is prepared to receive inbound requests, not only the response to the request, on the same connection. When that happens, the server and client sets up a connection table where the content of the certificates are stored – the domains and host names. Now if one of them has a request that is targeted to the same domain and the same IP and port (after DNS SRV lookups), the connection can be reused.

Checking and caching the certificate content

When the connection is initiated, both ends provide TLS certificates that contain one or multiple names or SIP uri’s. This list is cached and associated with the session. Now, if the same server host multiple domains, you can’t use the connection if the domain that is the target of the request doesn’t match the names in the certificate. In that case, you need to open a new connection to the same server.

Use DNS host names instead of IP addresses

On a related topic, notice that the RFC always use host names and not IP addresses in the Via: headers. This is of course a requirement if you want to match certificates. For TLS to work in all directions, host names should be used in Via: and Record-Route/Route headers. With a GRUU, you can also have a domain in the Contact. This also helps IPv4/IPv6 dual stack handling, letting every path select the optimal connection.

Combined with SIP outbound we have open connections all the way

Connection reuse is an important feature for all SIP servers, B2BUAs like Asterisk and SIP servers like Kamailio. Without it TLS will be hard to use and cause delays that will affect the calls. In combination with SIP Outbound, where the UA manages the connections to the first-hop servers, it is a working solution for TLS over NAT as well. Keeping TCP/TLS connections open like this is not new, Jabber/XMPP has done this from start. It’s just new to SIP.

I think SIP Connection Reuse support should be on the list of requirements when you select your next SIP application server for your Open Unifed Communication platform.

A modern SIP invite with SIP identity, outbound, ice and hopefully S/MIME will cause UDP fragmentation. A SIP stack that wants to stay up to date supports TCP. Anyone implementing a SIP network wants to support TCP. When we are there, beyond UDP, we might as well think about using TLS. Is it time to produce a SIP profile for user agents? SIP UA 2011.

Lately, I’ve been going through a lot of SIP RFCs and drafts, trying to get an overview of the security suggested in all of these documents. The quality of this work, seen from a developer’s perspective, is quite poor. Sometimes it seems like authors think, “oh, we need to add that security stuff, so let’s add a few keywords like TLS and S/MIME here and there“. We need to get better in reviewing the drafts from a security perspective. Here are  some thoughts on instructions to RFC authors:

  • S/MIME: If you refer to S/MIME, make it very clear which certificates that are going to be used and how the certificate verification process should happen – which part of the SIP message should match with which part of the certificate? And which certificate should be used to encrypt?
  • TLS: If you refer to TLS, you need to be very clear on why – is this to provide authentication, confidentiality or something else? Does the solution require mutual authentication or just server authentication? If authentication is part of your solution, make it very clear how you verify the certificate with the message, down to SIP header fields and X.509v3/PKIX fields.
  • SIPS: If you suggest usage of SIPS, make it very clear on what this adds and how the message flow is supposed to look like. Is SIPS used in the request uri, the Contact or somewhere else? What is the effect? Make sure you really understand SIPS before this is added. Or even better, just avoid SIPS and let it fade away.
  • Certificate matching: If you refer to a certificate SubjAltName, make very clear if it’s a URI or a dnsName field that is required and preferences if there are multiple SubjAltNames in addition to the certificate subject.
The worst documenst so far are the RFCs related to SIP subscriptions. They suggest using S/MIME for encryption, but does not explain how. Now, if I subscribe to the presence status of sip:bob@example.com, my SUBSCRIBE request will end up at the presence server for the domain example.com. Should the user agent somehow find the certificate for sip:example.com to encrypt the message? Should we use the certificate of sip:bob@example.com – which would require the presence server to have the private key belonging to Bob? The RFCs doesn’t help at all.
RFC 3857 states the following on the topic of eavesdropping on SUBSCRIBE/NOTIFY requests:

“To prevent that, watchers MAY use the sips URI scheme when subscribing to a watcherinfo resource.  Notifiers for watcherinfo MUST support TLS and sips as if they were a proxy (see Section 26.3.1 of RFC 3261).”

This means that a UA should be able to SUBSCRIBE over a TLS connection, and get NOTIFY over – what? Remember that this was written before SIP Outbound was standardized. For a developer this means that the subscriber is required to have a TLS certificate and accept incoming connections on the TLS port if the Contact in the SUBSCRIBE is a SIPS uri. The RFC should discuss this in more detail.

Nine years after RFC 3261 we have a larger toolbox, including GRUUs, SIP Outbound, SIP Domain certificates, DNSsec and much more. It’s time we restart the work with a SIP security architecture and provide something that developers can implement and that users will clearly feel is a better and more trustworthy solution. The IETF mantra is “rough consensus and running code”. RFCs should make it easy to produce running code. The SIP RFCs fails do this on the topic of SIP security.





SIP over dual stacks - IPv4 and IPv6

Stay Connected - learn more about SIP & IPv6

Yesterday I found an Internet Draft called Testing Eyeball Happiness that gives examples on how to test dual stack deployments. There is a known issue with applications that retrieves multiple IP addresses from the same host name in DNS and , following current RFCs, test them sequentially with a preference for IPv6 addresses. The timeouts when things go bad with one flow are far longer than what the user accepts. Let’s say that Bob (you know him) use his SIP phone to place a call to Alice. Bob’s phone calls an outbound proxy, that wants to forward to another domain. This domain announces both IPv4 and IPv6 addresses in DNS for their proxy. Now, Bob’s proxy actually has an IPv6 address, but is not connected to the Internet with IPv6. The proxy will try connecting to Alice’s domain SIP proxy over IPv6 for quite a long time before it recognizes that there’s no connectivity. Hopefully it will then try another address, but the question is if the user is waiting for that to happen. In telephony, loosing seconds is a catastrophe, especially between requesting a call and getting the first ringing signal. Remember – this is not about media, this is only about signaling. Without signaling, we’ll never get into any media issues.

HTTP and Happy Eyeballs

We’ve seen this problem on the web. Browsers suddenly told us that large sites was not available. Turned out that the new home router enabled IPv6 tunnels and announced IPv6 prefixes on the LAN, something that the firewall blocked. By disabling IPv6 in the laptop, we could reach the web site again. This caused web sites to stop announcing IPv6 and computer owners to disable IPv6. This was no good for the IPv6 migration so the browser developers started to try to find solutions. The Happy Eyeballs discussion in the IETF is about finding algorithms where the browser connects to all addresses in parallel and selects a candidate that answers quickly. In SIP, we need to implement the same fix, over UDP, TCP and STCP. I’ll try to set up some tests at SIPit to see what the current state is.

A quote from the abstract section of the IETF draft:

In a dual stack network (i.e., one that contains both IPv4 [RFC0791] and IPv6 [RFC2460] prefixes and routes), or in an IPv6-only network that uses multiple prefixes allocated by upstream providers that implement BCP 38 Ingress Filtering [RFC2827], the fact that two hosts that need to communicate have addresses using the same architecture does not imply that the network has usable routes connecting them, or that those addresses are useful to the applications in question. In addition, the process of establishing a session using the Sockets API [RFC3493] is generally described in terms of obtaining a list of possible addresses for a peer (which will normally include both IPv4 and IPv6 addresses) using getaddrinfo() and trying them in sequence until one succeeds or all have failed. This naive algorithm, if implemented as described, has the side-effect of making the worst case delay in establishing a session far longer than human patience normally allows. This has the effect of discouraging users from enabling IPv6 in their equipment, or content providers from offering AAAA records for their services.


I’m currently swimming through the deep waters of SIP RFCs in order to get an overview of TLS implementation requirements. Reading RFC 3428 – The SIP Message Extension– I found something I did not know. In section 11, Security Considerations, the RFC states:

In normal usage, most SIP requests are used to setup and modify communication sessions. The actual communication between participants happens in the media sessions, not in the SIP requests themselves. The MESSAGE method changes this assumption; MESSAGE requests normally carry the actual communication between participants as payload. This implies that MESSAGE requests have a greater need for security than most other SIP requests. In particular, UAs that support the MESSAGE request MUST implement end-to-end authentication, body integrity, and body confidentiality mechanisms.

I have seen quite a few implementations of MESSAGE, but none has been compliant with RFC 3428.

The SIP MESSAGE implements a way to send short messages over SIP, within a dialog or outside of a dialog. MESSAGE requests does not create dialog, thus there’s no “session”. For chat sessions that , MSRP – the message session relay protocol – was developed. I’ll try to write more about that protocol in another blog post.

Having a lot of interesting discussions about Realtime Network Security, with SIP as a focus, these days. We need to get enough people in a large room with enough whiteboards to attack the issue. The SIP RFCs needs many updates in this area to help developers to develop more secure software.

Last week I talked at the Voip2Day conference in Madrid, organized by Avanzada7. The talk, named “Watch out!” covers new areas developed in SIP, but not implemented in many devices or servers out there. Solutions for NAT traversal, PSTN trunk registration and new work with the real time web is covered, along with a small update to the list of 10 bullets to remember when implementing a new SIP platform.

Some topics covered:

  • ICE, Interactive Connection Establishment, a complex but working solution to find a working media path between two Sip phones, either directly or using a media relay (A TURN server). Used both for NAT traversal and IPv4/IPv6 dual stack deployments.
  • SIP Outbound, the way to handle NAT traversal for SIP signaling. With SIP outbound, the client sets up multiple IP connections, called flows, to servers while indicating that it’s actually the same device that registers on all these connections. The proxy can then do failover if one connection fails. It’s up to the SIP phone, the user agent, to maintain the connections and re-open them when they fail.
  • GIN – the way SIPconnect sends a registration for a SIP trunk with multiple phone numbers. Before GIN, every vendor used it’s very own hack which raised the cost for service providers that wanted to support multiple vendors.
  • GRUU – Globally Routable User URI’s – a domain-based address for every device that registers for an account. Makes it possible to do more complex operations over domain boundaries. Without a GRUU, many URI’s are unusable since they’re referring to an IP address hidden behind a NAT device.
I feel that ICE and SIP outbound are good candidates on solving the NAT puzzle as well as the IPv6 transition. We need more Open Source implementations as a reference!
The presentation also covers RTCweb briefly. On the conference, there was a live demonstration by Iñaki Bas Castillo and a colleague of a SIP implementation in JavaScript connecting over WebSockets to a SIP proxy. They lacked RTCweb so there was no media in the calls, but it showed that it’s possible to implement SIP in the browser!

The talk is now published on Slideshare and can be viewed online. Enjoy!