Testing Shibboleth - Server and Network

In production, the network setup is controlled by Operations. In testing, the developer can run Shibboleth on a desktop Sandbox, or can "redirect" URLs from one host to another, or can use SSH tunnels to connect to VMs in the machine room. It is fairly easy to route a request to a selected Shibboleth server, but it can be fairly difficult to actually get it working. Shibboleth doesn't simply passively accept requests. Certain things get checked for correctness. Shibboleth can reject requests because of the way they arrived over the network, the contents of the Hosts HTTP header, and the use of https to transport the request.

Developers do not configure the F5 front end, and they generally just accept the VM that Operations gives them. So these topics are not necessarily part of a developer's technical training. This document explains a few of the complex options for transporting HTTP requests and the peculiar nature of SAML validation that can be affected by the network configuration. It enumerates the problems you may run into if you try to use a configuration other than the ones documented in the other Testing documents on this site.

IdP Initiated Login

An "IdP Initiated" test occurs when you click a URL that points to our Shibboleth server and provides the EntityID of one of our Service Provider partners:

http://localhost:8080/idp/profile/SAML2/Unsolicited/SSO?providerId=nobody.yale.edu

This URL calls a Shibboleth server that could be running under Tomcat on the local desktop or could be a test version of Shibboleth running on a VM in the machine room where you have created an SSH tunnel connecting port 8080 on your desktop to the real Server port on the real Server VM.

This URL is a Shibboleth feature, not part of the SAML standard. The Shibboleth code that responds to a "/profile/SAML2/Unsolicited/SSO" URL path does not care about the hostname, port, or protocol. It makes no assumptions about how the client connected to the server or what the network looks like.

The only checks are:

• the providerid ("nobody.yale.edu" in this example) must be the EntityID of a Service Provider configured in some Metadata file
• The Metadata file must have a configured default AssertionConsumerService URL
• The hostname in the ACS URL must exist on the network.

Internally Shibboleth builds a dummy SAML Request, then triggers normal SAML processing to generate a Response.

If Consent is configured (and it will be by default in most Sandbox environments) the generated Attributes will be displayed. Then the SAML Response will be sent to the ACS URL.

This will work and will attempt to deliver the Response no matter what operating system you are running on, no matter what hostname, port number, or protocol your Tomcat uses, no matter what EntityID your Shibboleth server is configured to use for itself. In short, there are no network issues for IdP Initated Login.

Of course, the SP will probably discard the Response because it will generally not be digitally signed correctly, and the EntityID that issued it may be unrecognized by the SP, but that is for later testing.

The Unicon CAS Integration

Unicon provides a Shibboleth Login ("Authn") module that uses CAS. To configure this integration, the properties file must supply the protocol and hostname to use in the Service string sent to CAS. CAS will Redirect the Browser back to this URL after it issues the Service Ticket. Unicon has a specific property name, but Yale used a different name long before we switched to using the Unicon code. We keep the old Yale property name and use it to set the new Unicon property.

So in the Install Project, you must set a property named "cas.target.url" to be the protocol and hostname of the Shibboleth server from the point of view of the Browser. When you are testing a Sandbox running on your desktop, then it makes perfect sense to set the property in install.properties and make the value to "localhost", or more specifically:

cas.target.url=http://localhost:8080

However, this is probably not a good value to use when Shibboleth is installed in DEV or TEST. In these cases the property is set in install-DEV or install-TEST and it will reference the "auth-dev.yale.edu" and "auth-test.yale.edu" machines. This means, however, that if you install Shibboleth into DEV or TEST normally with these parameters, then you are not going to be able to use these VMs through an SSH tunnel (where they appear to be localhost:8080) unless you use some mechanism (Redirectory, Charles Proxy, hosts table) to rename the "https://auth-dev.yale.edu/" in the CAS Redirect to actually go to the SSH Tunnel.

You know you have this problem when Shibboleth sends you to CAS, you login, and then get a

HTTP Status 500 - Error processing ShibCas authentication request

error with a root cause of:

ExternalAuthenticationException: No conversation state found in session for key (e1s1)

What has happened is that Shibboleth received your original request from what it regards as one browser session, and it sent that Browser to CAS, but now it is getting back a response from CAS that appears to be coming in from a different Browser (although it is really the same Browser connecting to Shibboleth through a different network path using a different host name).

This error can also occur when two Shibboleth VMs exist behind a load balancing front end, but the front end has not been configured to route subsequent requests for the same session (based on the "JSESSIONID" parameter) to the same VM. Then one VM sends the Browser to CAS, but the response gets routed to the other VM.

SP Initiated

In an SP Initiated login, the user goes to the application first and is then sent back to CAS. For example, if you go to "http://mail.bulldogs.yale.edu" to get your mail, then the Google Apps login through Shibboleth is SP Initiated. Other examples include yale.box.com, yale.service-now.com, and www.yale.edu/iptv.

The Service Provider generates a SAML Request message and sends it to Shibboleth. Unfortunately, the SP knows that Yale's Shibboleth server is located at "https://auth.yale.edu/idp/.." and not only does it tell the Browser to go to that network address, but it also includes this URL in the Request message. Shibboleth implements the requirement in the SAML standard that it must determine its own network URL and compare that string with the network URL sent in the SAML Request, and reject the SAML Request if the two don't match.

The problem here is that you cannot configure Shibboleth with a fixed URL string and tell it to use that string in the comparison. Instead, Shibboleth calculates its own URL using information in the HTTP headers, and information provided by Tomcat in the HTTPServletRequest object.

Of course, this is not a problem for Production Shibboleth because that machine really is "https://auth.yale.edu". The problem then is to try and trick a test version of Shibboleth running on another network address into believing that it really is on a host named "auth.yale.edu" and the request is coming in over SSL/TLS (https), on port 443, even if that is not the case.

If you don't get it right, then Shibboleth discards the Request, generates a error message in the log, and displays an error page. The log message will always say that it was comparing "https://auth.yale.edu/idp/..." to its own network address and the comparison did not match. Looking at the string it was trying to match, you will see what it generated:

• "http://localhost:8080/" -  You gave the Browser the local address and it passed this information on to Shibboleth
• "http://auth.yale.edu" - You generated the right hostname, but Tomcat has told Shibboleth this request came in on "http" instead of "https"
• "https://auth.yale.edu:8080" - Tomcat is willing to say the request was "https", but the real port number got reported

At one point I thought this problem was solvable. I now realize it is not worth the effort. Do all your basic testing using IdP Initiated URLs. Any final end-to-end testing, where you start with an SP URL and end with an actual login to you Eliapps mailbox or you Box files should be done using the Pre-Production machine behind the F5 by adding a line to the hosts file assigning the "auth.yale.edu" hostname to a test address.

Background on Proxies and Tunnels

In modern networks, the Browser almost never talks directly to a production server. Production servers are hidden away in machine rooms behind firewalls, and access to them goes through a network front-end device. At Yale, that device is called the "F5".

So in the simplest case, a Browser going to CAS, or Shibboleth, or any other production service actually connects to the F5. The F5 then forwards the Browser request into the machine room, and it sends back to the Browser the response from the real server.

In HTTP, when a computer stands between the Browser and the Web Server and acts as a silent intermediary, it is called a Proxy. There are actually two configurations. In the old days, the original Web Proxy was a device on the Yale Campus that held local copies of frequently used pages from distant Internet servers. That way hundreds of requests from Yale users for the same front page of the New York Times could all be satisfied by a single copy saved from a recent request. That was useful back when the Internet was slow, but today the Internet is so fast and powerful that this type of proxy has become obsolete.

So mostly today we talk about a Proxy that sits in front of a server, rather than a Proxy that sits behind the Browser. This is often called a "Reverse Proxy" and it receives requests intended for a Web Server and, in many cases, it distributes the requests that come in among a number of identical server computers to spread the work around and quickly recover if one of the servers has a problem.

The Proxy is supposed to be invisible. To the Browser, the proxy appears to be "auth.yale.edu" or whatever hostname the Browser put into the URL. If that is the case, the actual VM in the machine room where a service like Shibboleth is running will never really be named "auth.yale.edu" but will instead be "vm-shibprd-01.web.yale.internal". Furthermore, that real machine name and the IP address of that machine will not be visible outside the machine room and you can only get to it through the F5.

So you can see that if Shibboleth needs to know its own URL from the point of view of the Browser, and it insists on figuring it out for itself through programming, there is a bit of a problem. The "auth.yale.edu" name is known to the F5, but until Shibboleth gets a request from the first user, the name "auth.yale.edu" is configured nowhere on the VM and there is nothing Shibboleth can do to find it.

In HTTP the Browser puts the protocol and host name in a Host: header that is part of the data sent to the server. So coming out of your Browser, there will be a line containing "Host: https://auth.yale.edu". If that line gets to Shibboleth unchanged, then Shibboleth can use it to guess the right answer to the "What is my network name" question. Of course, the F5 or any other proxy can change this header, but then there are other network standards that have developed over time to create other Headers that track such changes.

Tomcat was written to work correctly in the modern world of firewalls, and reverse proxies, and F5 devices. It can be configured to report to the application that the hostname is "auth.yale.edu", that the data came in over "https" on port number 443, despite the fact that the data actually came in on a "http://localhost:8080" request.

Feel free to download the Shibboleth source code and read it. You will discover that there is some Shibboleth source that calls some OpenSAML source that calls back to some other Shibboleth source. Some information comes from Tomcat, some from Spring, and unfortunately when you are talking about how the CAS integration does the same thing then Shibboleth calls the Unicon code which in turn calls the CAS Client code.

Somehow under the covers this stuff digs through the HTTP headers and the Tomcat configuration and comes up with what it thinks is "The URL the Browser used or could use to communciate to this Shibboleth server". Most of the time it is right, or when it is wrong it is obvious why it is wrong. If you cannot fix it in 10 minutes, you will never fix it.

That is why the simplest solution is to put your code in Pre-Production and change the hosts table to point "auth.yale.edu" to the F5 address that routes to Pre-Production.