Added the metadata url as an attribute on the IDP in the keycloak saml configuration which then propagates through to the DefaultSamlDeployment where it is used on the construction of the SamlDescriptorPublicKeyLocator thereby allowing support for ADFS or other IDP which uses a path that is different to the Keycloak IDP.
To make this work when testing with ADFS a change was made to SamlDescriptorIDPKeysExtractor because it would not extract keys from metadata which contained the EntityDescriptor as the root element. The solution was to change the xpath expression in SamlDescriptorIDPKeysExtractor so that it does not require a wrapping EntitiesDescriptor but instead loads all EntityDescriptors located in the document. This allows it to handle a single EntityDescriptor or multiple descriptors wrapped in an EntitiesDescriptor in the same xpath expression. A unit test SamlDescriptorIDPKeysExtractorTest has been added which validates that keys can be loaded in both scenarios.
This is an issue with the Spring Security Keycloak Adapter relating to
the way the Authentication is stored in the SecurityContext, causing a
race condition in application code using that. It does not seem to
affect actual Spring Security operation.
We had a pretty strange race condition in our application. When many
requests were incoming at the same time, occasionally the old
unauthenticated Authentication provided to
KeycloakAuthenticationProvider for performing the actual authentication
would stay the current authentication, as returned by
SecurityContextHolder.getContext().getAuthentication(). That resulted
in authenticated users' JavaScript requests occasionally (~1/50 given a
large request volume) returning a 403 because the 'old' token was still
in the context, causing Spring Security to see them as unauthenticated.
This PR resolves this issue by replacing the whole context, as suggested
by a Spring Security contributor in jzheaux/spring-security-oauth2-resource-server#48. By default,
SecurityContextHolder keeps the actual context object in a ThreadLocal,
which should be safe from race-conditions. The actual Authentication
object, however, is kept in a mere field, hence the reason for this PR.
JIRA issue: https://issues.jboss.org/browse/KEYCLOAK-9539
The root cause is that NodesRegistrationManagement.tryRegister can be
called from multiple threads on the same node, so it can require
registration of the same node multiple times. Hence once it turns to
tasks that invoke sendRegistrationEvent (called sequentially), the same
check has been added to that method to prevent multiple invocations on
server side, or invocation upon undeployment/termination.
