Usually in SAML and most similar SSO protocols, the flow is assumed to be a service requesting authentication by redirecting the client to the login service, and then getting back a response. In the original SAML 1.0 and SAML 1.1 standards, though, SSO was described in only semi-interoperable terms as a "pushed" response from the IdP to the SP, and the "request" portion was left out. This was carried into SAML 2.0 as a mode called "IdP-initiated" or "unsolicited" SSO.
While this approach lacks interoperability, it has perceived benefits for some service providers; they get to do less work and push that work onto users and IdPs. So it isn't unusual to find SPs that refuse to support the standard fully and insist on this approach.
What is misunderstood about this feature is that it is not interoperable, despite being part of the standard. Interoperability requires a well-defined message, and the basic idea behind IdP-initiated SSO is that the message is up to the IdP. Something has to initiate the process, it can't magically start for no reason. So there is a request to the IdP, but it isn't a SAML-defined message and no two IdPs are likely to work the same way. It's a proprietary mechanism.
There is no special configuration for this use case, it's subsumed into supporting Browser SSO for SAML 1 and SAML 2 for a relying party. In the default configuration, SAML 2.0 is enabled using the profile configuration bean named "SAML2.SSO", the same one that configures the behavior of the normal request/response usage of the profile.
Note that while SAML metadata for the SP is required, as it always is, there's nothing in particular about this use case that's evident from the metadata you have to provide (or, often, build yourself; the Venn diagram of "requires IdP-initiated SSO" and "doesn't supply metadata" is round). The SP's metadata is in general the same as it would be in any other scenario.
There is currently no simple method of disabling the Unsolicited SSO support for SAML 2 separately from the overall support for SAML 2 SSO. Blocking the URL outright via a web server rule is the most obvious way to do it.
One way you can disable support for this feature for specific services ungracefully is by modifying their SAML metadata to include AuthnRequestSigned="true" in the <SPSSODescriptor> element. Doing so causes the IdP to require requests from that SP to be signed, and since this protocol does not allow for signing, it will cause such requests to fail with an error.
As described above, the protocol for this feature is proprietary and is defined by the Shibboleth software. The interface to ask the IdP to respond to an SP without the SP having made a request involves the use of some defined query string parameters.
The examples assume the default locations supported out of the box, which should rarely need to be adjusted, and an IdP located at idp.example.org.
They also assume proper metadata is loaded into the IdP. There is nothing special about the metadata required to use this feature, it's the same metadata required for any use of the corresponding SP, and if you have to create that by hand because you're working with a deficient partner and/or outside the context of a federation, see other topics and examples to explore how to do that.
Given an SP named "https://sp.example.org/shibboleth", requesting SAML 2.0 SSO to the SP's default endpoint in metadata is just a link to:
The example above is something you'll run into occasionally: the doubly-encoded value. When an SP is sufficiently broken, it may require query parameters that are themselves passed as values of the target parameter (as shown above, where the target value is a query parameter named "rpid" and with the value of another URL). The URL parameter to which the "rpid" name is assigned is URL-encoded, and then the entire name/value pair is URL-encoded for inclusion in the "target" parameter.
It's not common to have to provide the "shire" parameter in the SAML 2.0 case, but it may be needed if a single SP has development and production versions distinguished by SAML endpoint; the parameter would tell the IdP where to send the browser after logging in, and thus which environment to access.