MetadataDrivenConfiguration

Overview

One of the challenges of dealing with interoperability issues with SPs is maintaining all the custom configuration rules needed to drive the IdP's behavior in all kinds of different ways. Traditionally most of these kinds of settings were concentrated in the relying-party.xml file and involve defining "overrides" for SPs or groups of SPs, often through manually maintained lists of entityIDs.

As discussed in the RelyingPartyConfiguration topic in the section on dynamically-driven profile configuration, the IdP supports the ability to produce these settings on the fly at runtime, which creates opportunities for more flexible configuration, at some cost in complexity.

The best way to leverage the features described in that section involves the use of metadata "tags" (a shorthand name for the <mdattr:EntityAttributes> SAML metadata extension) to embed signals in metadata that can drive configuration. In the longer term, any hope for a simplified approach to configuration or even a GUI rests on a strategy of this general type.

The IdP supports a built-in layer of code and Spring wiring to enable the use of the metadata-driven approach to configuration in a systematic way that provides a set of conventions for tag names and values to use to drive a significant range of configuration settings. With a few up-front adjustments to the relying-party.xml file, it's possible to enable comprehensive support for metadata-driven configuration that can be freely intermixed in most cases with more static settings as required.

General Configuration

Support for this feature has been added to the runtime environment in a way that avoids overhead (and risk) if not used but allows for straightforward enablement on a per-profile basis. The existing parent beans used for defining relying party defaults, overrides, and profile configurations all have analogs with a ".MDDriven" suffix in their names. Changing the existing parent or referenced bean name by adding the suffix and reloading the configuration will cause the runtime to evaluate most of the commonly-used configuration settings by checking for appropriately-named "tags" in the SP's metadata before falling back to the default, or statically configured, value if no tag is found.

For example, if you would like to enable tag-driven configuration for SAML 2.0 Browser SSO, you can replace any reference to the SAML2.SSO bean with SAML2.SSO.MDDriven, as in this example:

Attribute / Property Convention

The "bridge" between the metadata and this feature is the naming and syntax of the attribute tags used to control the configuration properties in the IdP. As an implementation strategy this could be "brute forced" with arbitrary mapping dictionaries, but the supplied implementation takes the approach of automating a mapping between the names of settings and the identifiers used internally for the various profiles, and the names of the corresponding SAML Attributes in metadata for those profile/setting combinations.

Each setting has a "base" name matching its Java bean property name as would be used if the setting were explicitly configured in Spring. For example, the property set via p:defaultAuthenticationMethods is named "defaultAuthenticationMethods".

The corresponding SAML Attribute for a setting is named by suffixing the "base" name of the setting to a profile URL that is defined by the Shibboleth software for each of the supported profiles.

The URLs are as follows:

It follows that the includeAttributeStatement property of the "Shibboleth.SSO" profile configuration can be set via a metadata Attribute named "http://shibboleth.net/ns/profiles/saml1/sso/browser/includeAttributeStatement"

As an additional convention, a setting can be configured for all profiles simultaneously by prefixing it with the URL "http://shibboleth.net/ns/profiles"

Applying Tags

There are two ways to apply tags to metadata: directly and indirectly.

The direct method applies to cases in which you control the metadata or when the source of the metadata supports the inclusion of the necessary extensions. The tags simply appear within the metadata being loaded into the IdP using an <mdattr:EntityAttributes> extension element:

The indirect method uses the EntityAttributesFilter to attach the tags at runtime after loading and verifying the metadata. This is useful for cases in which you want to rely on externally supplied metadata but still organize parts of your configuration around the metadata. One use for this is simply consistency: if you have both external and local metadata, you can drive the configuration with both.

Relying Party Configuration Examples

The following examples illustrate possible <Attribute> syntaxes that can be used to configure a variety of common features. The examples apply equally to either the direct or indirect methods of attaching the attributes to the metadata.

Signed Assertions vs. Responses

Enabling signed assertions for a particular SP is advisedly handled by setting the WantAssertionsSigned XML attribute in metadata, but isn't always possible and sometimes has to be combined with disabling signed responses (or just for efficiency).

Disabling Encryption

Setting idp.encryption.optional is usually a workaround for handling the majority of SPs without encryption support, but there are a couple of scenarios in which it's useful to be able to manually disable it. Some federations (InCommon for one, at least for the time being) have limitations such that SPs without encryption support are stuck registering keys they don't support. Some SPs support encryption but build in time-bombs by forcing flag day key rotations on all IdPs that cause outages or manual work. Or you may simply want to drive this more explicitly for only a subset of SPs and not as a global change.

Note that it may be helpful to make sure encryption also doesn't get used during logout, so the second tag may be useful.

NameID Format Selection

Using <NameIDFormat> elements in metadata (which can also be added at runtime with the NameIDFormatFilter) is the normal way to use metadata to select the Format to use, but the "unspecified" Format has to be triggered with a profile setting. That's not commonly needed, but it's easy to define a tag for in such cases (and if you prefer you can obviously designate any Format this way).

One concern is that any two SPs using it are only coincidentally going to want the same data, so this isn't always solely a matter of Format selection. In such cases, it may be necessary to add activationCondition properties based on SP to specific generators that are handling the same Format.

Bug Workarounds

Some of the other profile settings are workarounds for bugs, e.g., omitting the NotBefore attribute. This is not very common but easily tag-driven.

A more obscure example would be an SP that requires additional <Audience> values in addition to its own entityID. This mainly demonstrates that some settings may be multi-valued.

Forcing MFA

Handling SPs that require MFA but can't request it requires IdP-side configuration, usually involving a couple of settings. The URL below is just an example, it has to be replaced by whatever <AuthnContextClassRef> constant is used in a deployment to signal MFA, or possibly more than one.

Interceptor Flows

Triggering consent based on the SP is pretty common.

The authorization intereptor flow is another case, though the checking logic would have to be extended for each different service/scenario. In the example, both flows are enabled.

Legacy Profiles

A common use case is enabling SAML 1 for legacy systems, often combined with either enabling queries or attribute push (to eliminate the queries). This is not handled as well via this mechanism because one generally enables most profiles by default, and it's not practical to enable them and then try and disable them for all but a few SPs. This goal is the opposite, disable by default and enable the profiles for a few exceptions.

Therefore, the built-in wiring cannot accomodate this use case, but it's possible to do this with an activation condition. This is the same wiring that is provided by the system configuration for many of the other properties but is supplied here "by hand" in order to adjust the default value of the condition to "false" instead of "true".

Attribute Filtering

The AttributeFilterConfiguration has had support for metadata-driven configuration for a while now, but it hasn't been extensively used. A logical approach is to align usage with the property-driven model outlined above and move to a per-attribute policy model, in contrast to the fairly common model today of defining policies around services. Since the filter layer operates by iterating over all policies to determine if they apply, it may be more efficient for larger policy sets to redesign policies around tags that signal release of each individual attribute.

To facilitate the sharing of examples, building of tools, and a more useful set of default rules in the software, we included examples in the new default filter policy file using the following tag:

There’s nothing special about this Attribute name/convention, but we needed to pick something we controlled for demonstration’s sake.

Example

Here's an example policy (more or less matching an example in the default file) that applies this tag test to a couple of attributes. A couple of subtle points here.

One is that this approach is really a purely "local" one because the IdP Attribute names are local/internal only. You might assume most people use similar conventions and you'd be right, but you can't expect that to be true universally, so it is not appropriate to ever try and use this kind of metadata tag outside of a scenario that does not involve control of the IdP. It's not a fix for the brokenness of the more general SAML <RequestedAttribute> metadata element, which is usually impractical for a variety of reasons.

Less critically, note that this policy example demonstrates an optimization in that it applies the tag check in the <AttributeRule>(s) and not up in the <PolicyRequirementRule>. This has the advantage of requiring less XML to express and is about the same or better in performance than running multiple policies. It's not as good for an attribute that might have 3-5 or more values, since that would force the cost of checking for the tag to be paid for each value.

Reference

Beans