The IdP software uses the Apache HttpClient library more or less anywhere that this functionality is required, which for most deployers is confined to obtaining metadata from remote sources. The necessary settings to control the behavior of the metadata client code can be handled directly in the metadata-providers.xml file in most cases, so this topic is primarily a reference for people who have very advanced needs or are using other components and features that make use of HTTP clients.
Some of the components that require or at least support the injection of a custom client bean include:
the HTTPConnector for web service access in the attribute resolver
You might need to dive into this topic further if you want to finely tune settings for different situations or if you have advanced security requirements that go beyond default behavior. A common example would be if you want to control TLS server authentication at a finely-grained level to avoid dependence on the default trust behavior of Java's TLS implementation. This is particularly true if you ever find yourself modifying the "global" Java trust store. That should never be done, since it makes Java upgrades much more troublesome.
In comparison to some of the IdP's features, the veneer here is very "thin". That is, we don't have a lot of layers of abstraction and simplification in place to hide the gory details (this may come as news, but much of the rest of the configuration is very abstracted and simplified from what it could look like).
As a result, this topic is heavy on examples, and going beyond the examples is more a case of reading javadocs and finding the right settings than reading some documentation to magically impart the right answer.
A set of Spring factory beans are provided that understand how to build an HttpClient instance with a variety of features and settings. For technical reasons the basic caching behavior of the client is determined by selecting from among three different factory bean types:
As you would expect, the first provides no explicit caching of results, the second caches results on disk (but not across restarts of the software), and the third caches results in memory. This is HTTP caching; that is, it relies on signaling between the client and web server to detect when to reuse results and supports conditional GET requests with cache control headers to redirect requests into the cache. Essentially they act much like a browser would.
In the vast majority of cases, we strongly recommend sticking with the non-caching one (the first one). The other two don't really provide benefits in most use cases and they tend to behave unexpectedly and they interfere with commonly-used components that already provide HTTP caching features or file backups for resiliency.
There were a set of built-in client beans defined prior to V4 that exposed a bug in Spring that remains unfixed; as such, we strongly advise NOT relying on the following deprecated beans when defining your own clients:
We instead suggest relying on a new set of parent beans (the link will take you to the relevant javadoc):
However, be advised that these beans merely simplify the declaration of the underlying class, and do NOT predefine any properties or non-default behavior at all. You must define any settings you wish to use yourself. You may wish to copy the deprecated bean declarations above from the underlying system resources for your own use, but you should use the new parent beans and define separate clients for any use case for which you want to customize the behavior. (Sharing a single client bean across different components generally can work, but when this happens across different reloadable services, an unfixed Spring bug becomes a problem.)
Many common settings are available as properties (mostly commented out for easy definition in services.properties).These properties are a mix of global settings that configure each client type and specific settings controlling the caching behavior of the individual types. This is a suitable approach if you have one use case, or are fine with all use cases sharing client behavior, and your own beans may chose to reference those properties if you prefer that approach.
A common use for advanced configuration is to control the server certificate evaluation process. While the IdP typically relies on SAML metadata for this when communicating with servers, this doesn't work for a lot of non-SAML use cases. However, much of the same machinery can be repurposed to support a variety of trust models, and we have integrated this into the HttpClient library. Another use case is to support HTTP authentication such as a name and password to access a resource.
By default, the HttpClient will rely on Java's built-in TLS behavior and code built into the HttpClient library, and perform basic hostname/certificate checks and will rely on the Java global trust store to determine whether to trust the certificate (or its issuer). This is where most non-Shibboleth software stops. In testing scenarios, you can turn off this checking via the connectionDisregardTLSCertificate property, but this should never be used in production.
For more advanced control over this process, you need to perform additional steps, and add more complex Spring wiring involving the org.opensaml.security.httpclient.HttpClientSecurityParameters class, which provides for using a TrustEngine to evaluate the server's certificate. There are a variety of TrustEngine implementations that can perform simple or advanced checks, but the critical difference is that they're contextual. That is, they can be applied to one component such that the rules it uses can be specific to that component alone and not the whole system.
A particularly useful approach is to abandon the fundamentally flawed use of commercial certificates with non-browser-facing services and use self-signed certificates evaluated on the basis of public key comparisons, much like Shibboleth does with SAML, or even certificates issued on an automated short-term basis by dedicated CAs.
If you want an HttpClient bean to use the special TLSSocketFactory we provide that supports a TrustEngine, you MUST provide an HttpClientSecurityParameters instance to the component using the HttpClient bean to configure the security behavior you want.
Repeating: you can tell the HttpClient bean that you want to support a more "advanced mode" of processing, but you configure the rules not on the HttpClient bean but rather on the component using the client bean. This is because the injection of the rules you want to apply have to be added at runtime when the client gets used and not into the client's own data structures. It's a consequence of the library's design.
To use a TrustEngine, you need to define an HttpClientSecurityParameters bean with a tLSTrustEngine property. While you can define any compatible object, the two most common cases are supported via a pair of factory beans:
Explicitly compares a set of trusted public keys against the key in the server's certificate, with no regard for the rest of the certificate.
Provides a full suite of PKIX processing to the server's certificate, but with dedicated trust authorities and policy rules.
The examples below are semi-complete, in that they include a portion of the "real" component that the security rules are being supplied for. Most of the components that can support this will carry both httpClient and httpClientSecurityParameters properties. Components defined in custom (non-Spring) XML syntax will usually support httpClientRef and httpClientSecurityParametersRef XML attributes to reference Spring beans of the proper type.
Unsurprisingly, the "explicit" engine is a bit simpler to use. You can provide any number of public keys and certificates via resources (file, classpath, even HTTP, though that gets pretty circular here), to drive the engine.
The common case of a rule to compare the key against a certificate loaded from a file looks like this:
The full range of PKIX options is quite complex, but for basic use cases a factory bean makes it simple. To validate the server's certificate against a fixed CA (name checking is turned off because the HttpClient is already doing this step for us):
Other PKIX options include the verifyDepth property to control the chain length, and the cRLs property to supply certificate revocation lists.
Applying a TrustEngine to HttpClient
You should start by configuring a component using the HttpClient with the HttpClientSecurityParameters wiring needed to implement your needs, as in the above examples.
Once that's in place, temporarily configure the HttpClient with the connectionDisregardTLSCertificate flag on, and finally just test it and let it fail and log something like the following:
19:04:53.364 - 127.0.0.1 - WARN [org.opensaml.security.httpclient.HttpClientSecuritySupport:98] - Configured TLS trust engine was not used to verify server TLS credential, the appropriate socket factory was likely not configured
19:04:53.366 - 127.0.0.1 - ERROR [net.shibboleth.idp.profile.impl.ResolveAttributes:299] - Profile Action ResolveAttributes: Error resolving attributes
net.shibboleth.idp.attribute.resolver.ResolutionException: Data Connector 'webservice': HTTP request failed
Caused by: javax.net.ssl.SSLPeerUnverifiedException: Evaluation of server TLS credential with configured TrustEngine was not performed
The category and details in the ERROR will vary by component, but the WARN and SSLPeerUnverifiedException messages will consistently indicate that the component was configured to apply a TrustEngine to the connection but couldn't do so. This may happen after the connection is established, possibly even after data is sent and consumed, but it should happen at the end. If not, something is wrong and has to be corrected to be confident that the eventual configuration will work.
Once the warning/error occurs, only then make the final change to the HttpClient bean to fix the error by overriding the tLSSocketFactory property (note the weird mixed case, that's due to the usual "first character is lower case" property convention in Java beans). The IdP includes a pair of socket factory beans for this purpose, shibboleth.SecurityEnhancedTLSSocketFactory and shibboleth.SecurityEnhancedTLSSocketFactoryWithClientTLS
So the final step to adding advanced TLS support is, for example:
<!-- Include other properties as desired... -->
<bean id="SecurityEnhancedHttpClient" parent="shibboleth.HttpClientFactory"
You won't generally need to configure your own instance of the socket factories, because all the interesting settings are part of the other component's configuration, via HttpClientSecurityParameters.
The code as it stands does not generally support TLS Renegotiation, which is most commonly encountered when using a virtual host that applies client TLS to only a subset of paths and not the host as a whole.
Configuring a component using the HttpClient with a private key and certificate for authenticating itself to a server is a simple two step process:
Make sure the HttpClient bean's tLSSocketFactory-ref property points to the shibboleth.SecurityEnhancedTLSSocketFactoryWithClientTLS bean.
The syntax for supplying a keypair can be essentially copied from the credentials.xml file that contains the more "usual" keys and certificates used by the IdP. Note that the beans defined in that file are not visible outside the RelyingPartyConfiguration so if you try to reuse them elsewhere by reference, you'll get errors. The example builds on the previous one and includes server authentication.
Currently it is less than straightforward to configure more typical HTTP credentials such as a basic-auth username and password, due to a lack of abstraction methods in our code to hide some of the gory details of the HttpClient's data model. In particular, some of the methods that need to be called take multiple parameters, which violates the bean convention for a setter. It's possible to invoke more complex methods in Spring, but it takes some extra wiring. We intend to supply some additional code for this in a future release.
Note that some of the older custom schemas such as the metadata configuration schema may support shorthand for supplying username/password credentials, and while these do work, they're deprecated in favor of the more generic httpClientSecurityParameters-ref syntax.
At the moment, it's fairly simple to supply a username and password that gets used unilaterally with a given component's requests. That is, it's not "scoped" to limit its use to a particular server. This implies that you have a working configuration in place to authenticate the server's certificate so that the password isn't sent inadvertently to a malicious location. An example follows (again, building on the server authentication case):
The next level up in complexity is the desirable ability to limit the scope of the credentials for safety's sake. The example relies on the hostname and port of the server to scope the password. There are more advanced ways to build the AuthScope object being passed into the API such as including the Realm challenge from the server.
A more advanced example would be to configure multiple sets of credentials for different servers, assuming a component that potentially contacts different servers. Since this is not a common case with any of our components, it's not likely to be needed much.
These properties are used in a set of DEPRECATED parent beans that are no longer supported due to a Spring bug that can impact the reloading of service configurations. They remain present for compatibility and for convenience should you choose to use them in your own bean definitions