GEP-1897: BackendTLSPolicy - Explicit Backend TLS Connection Configuration

• Issue: #1897
• Status: Experimental

TLDR

This document specifically addresses the topic of conveying HTTPS from the Gateway dataplane to the backend (backend TLS termination), and intends to satisfy the single use case “As a client implementation of Gateway API, I need to know how to connect to a backend pod that has its own certificate”. TLS configuration can be a nebulous topic, so in order to drive resolution this GEP focuses only on this single piece of functionality.

Furthermore, for Gateway API to handle the case where the service or backend owner is doing their own TLS, and the service or backend owner wants to validate the clients connecting to it, two things need to happen:

• The service or backend owner has to provide a method for the Gateway owner to retrieve a certificate.
• Gateway API has to provide a way for the Gateway to configure and apply the validation options.

Immediate Goals

1. The solution must satisfy the following use case: the backend pod has its own certificate and the gateway implementation client needs to know how to connect to the backend pod. (Use case #4 in Gateway API TLS Use Cases)
2. In terms of the Gateway API personas, only the application developer persona applies in this solution. The application developer should control the gateway to backend TLS settings, not the cluster operator, as requiring a cluster operator to manage certificate renewals and revocations would be extremely cumbersome.
3. The solution should consider client certificate settings used in the TLS handshake from Gateway to backend, such as server name indication, trusted certificates, and CA certificates.

Longer Term Goals

These are worthy goals, but deserve a different GEP for proper attention. This GEP is concerned entirely with the controlplane, i.e. the hop between gateway and backend.

1. TCPRoute and GRPCRoute use cases are not addressed here, because at this point in time these two route types are not graduated to beta.
2. Mutual TLS (mTLS) use cases are intentionally out of scope for this GEP for two reasons. First, the design of Gateway API is backend-attached and does not currently support mutual authentication, and also because this GEP does not address the case where connections to TLS are implicitly configured on behalf of the user, which is the norm for mTLS. This GEP is about the case where an application developer needs to explicitly express that they expect TLS when there is no automatic, implicit configuration available.
3. Service mesh use cases are not addressed here because this GEP is specifically concerned with the connection between Gateways and Backends, not Service to Service. Service mesh use cases should ignore the design components described in this proposal.

Non-Goals

These are worthy goals, but will not be covered by this GEP.

1. Changes to the existing mechanisms for edge or passthrough TLS termination
2. Providing a mechanism to decorate multiple route instances
3. TLSRoute use cases
4. UDPRoute use cases
5. Controlling TLS versions or cipher suites used in TLS handshakes. (Use case #5 in Gateway API TLS Use Cases)
6. Controlling certificates used by more than one workload (#6 in Gateway API TLS Use Cases)
7. Client certificate settings used in TLS from external clients to the Listener (#7 in Gateway API TLS Use Cases)
8. Providing a mechanism for the cluster operator to override gateway to backend TLS settings.

Already Solved TLS Use Cases

These are worthy goals that are already solved and thus will not be modified by the implementation.

1. Termination of TLS for HTTP routing (#1 in Gateway API TLS Use Cases)
2. HTTPS passthrough use cases (#2 in Gateway API TLS Use Cases)
3. Termination of TLS for non-HTTP TCP streams (#3 in Gateway API TLS Use Cases)

Overview - what do we want to do?

Given that the current ingress solution specifies edge TLS termination (from the client to the gateway), and how to handle passthrough TLS (from the client to the backend pod), this proposed ingress solution specifies TLS origination to the backend (from the gateway to the backend pod). As mentioned, this solution satisfies the use case in which the backend pod has its own certificate and the gateway client needs to know how to connect to the backend pod.

Gateway API is missing a mechanism for separately providing the details for the backend TLS handshake, including (but not limited to):

• intent to use TLS on the backend hop
• client certificate of the gateway
• system certificates to use in the absence of client certificates

Purpose - why do we want to do this?

This proposal is very tightly scoped because we have tried and failed to address this well-known gap in the API specification. The lack of support for this fundamental concept is holding back Gateway API adoption by users that require a solution to the use case. One of the recurring themes that has held up the prior art has been interest related to service mesh, and as such this proposal focuses explicitly on the ingress use case in the initial round. Another reason for the tight scope is that we have been too focused on a generic representation of everything that TLS can do, which covers too much ground to address in a single GEP.

The history of backend TLS

Work on this topic has spanned over three years, as documented in our repositories and other references, and summarized below.

In January 2020, in issue TLS Termination Policy #52, this use case was discussed. The discussion ended after being diverted by KEP: Adding AppProtocol to Services and Endpoints #1422, which was implemented and later reverted.

In February 2020, HTTPRoute: Add Reencrypt #81 added the dataplane feature as “reencrypt”, but it went stale and was closed in favor of the work done in the next paragraph, which unfortunately didn’t implement the backend TLS termination feature.

In August 2020, it resurfaced with a comment on this pull request: tls: introduce mode and sni to cert matching behavior. The backend TLS termination feature was deferred at that time. Other TLS discussion was documented in [SIG-NETWORK] TLS config in service-apis , a list of TLS features that had been collected in June 2020, itself based on spreadsheet Service API: TLS related issues.

In December 2021, this was discussed as a beta blocker in issue Docs mentions Reencrypt for HTTPRoute and TLSRoute is available #968.

A March 2022 issue documents another request for it: Provide a way to configure TLS from a Gateway to Backends #1067

A June 2022 issue documents a documentation issue related to it: Unclear how to specify upstream (webserver) HTTP protocol #1244

A July 2022 discussion Specify Re-encrypt TLS Termination (i.e., Upstream TLS) #1285 collected most of the historical context preceding the backend TLS termination feature, with the intention of collecting evidence that this feature is still unresolved. This was followed by GEP: Describe Backend Properties #1282.

In August 2022, Add Provisional GEP-1282 document #1333 was created, and in October 2022, a GEP update with proposed implementation GEP-1282 Backend Properties - Update implementation #1430 was followed by intense discussion and closed in favor of a downsize in scope.

In January 2023 we closed GEP-1282 and began a new discussion on enumerating TLS use cases in Gateway API TLS Use Cases, for the purposes of a clear definition and separation of concerns. This GEP is the outcome of the TLS use case #4 in Gateway API TLS Use Cases as mentioned in the Immediate Goals section above.

API

To allow the gateway client to know how to connect to the backend pod, when the backend pod has its own certificate, we implement a metaresource named BackendTLSPolicy, that was previously introduced with the name TLSConnectionPolicy as a hypothetical Direct Policy Attachment example in GEP-713: Metaresources and PolicyAttachment. Because naming is hard, a new name may be substituted without blocking acceptance of the content of the API change.

The selection of the applicable Gateway API persona is important in the design of BackendTLSPolicy, because it provides a way to explicitly describe the expectations of the connection to the application. BackendTLSPolicy is configured by the application developer Gateway API persona to signal what the application developer expects in connections to the application, from a TLS perspective. Only the application developer can know what the application expects, so it is important that this configuration be managed by that persona.

During the course of discussion of this proposal, we did consider allowing the cluster operator persona to have some access to Gateway cert validation, but as mentioned, BackendTLSPolicy is used primarily to signal what the application developer expects in the connection. Granting this expectation to any other role would blur the lines between role responsibilities, which compromises the role-oriented design principle of Gateway API. As mentioned in Non-goal #8, providing a mechanism for the cluster operator gateway role to override gateway to backend TLS settings is not covered by this proposal, but should be addressed in a future update. One idea is to use two types: ApplicationBackendTLSPolicy, and GatewayBackendTLSPolicy, where the application developer is responsible for the former, the cluster operator is responsible for the latter, and the cluster operator may configure whether certain settings may be overridden by application developers.

The BackendTLSPolicy must contain these configuration items to allow the Gateway to operate successfully as a TLS Client:

• An explicit signal that TLS should be used by this connection.
• A hostname the Gateway should use to connect to the backend.
• A reference to one or more certificates to use in the TLS handshake, signed by a CA or self-signed.
• An indication that system certificates may be used.

BackendTLSPolicy is defined as a Direct Policy Attachment without defaults or overrides, applied to a Service that accesses the backend in question, where the BackendTLSPolicy resides in the same namespace as the Service it is applied to. The BackendTLSPolicy and the Service must reside in the same namespace in order to prevent the complications involved with sharing trust across namespace boundaries. We chose the Service resource as a target, rather than the Route resource, so that we can reuse the same BackendTLSPolicy for all the different Routes that might point to this Service. For the use case where certificates are stored in their own namespace, users may create Secrets and use ReferenceGrants for a BackendTLSPolicy-to-Secret binding. Implementations must respect a ReferenceGrant for cross-namespace Secret sharing to BackendTLSPolicy, even if they don't for other cross-namespace sharing.

One of the areas of concern for this API is that we need to indicate how and when the API implementations should use the backend destination certificate authority. This solution proposes, as introduced in GEP-713, that the implementation should watch the connections to the specified TargetRefs (Services), and if a Service matches a BackendTLSPolicy, then assume the connection is TLS, and verify that the TargetRef’s certificate can be validated by the client (Gateway) using the provided certificates and hostname before the connection is made. On the question of how to signal that there was a failure in the certificate validation, this is left up to the implementation to return a response error that is appropriate, such as one of the HTTP error codes: 400 (Bad Request), 401 (Unauthorized), 403 (Forbidden), or other signal that makes the failure sufficiently clear to the requester without revealing too much about the transaction, based on established security requirements.

All policy resources must include TargetRefs with the fields specified here. In an upcoming extension to TargetRefs, policy resources may also choose to include SectionName and/or Port in the target reference following the same mechanics as ParentRef.

BackendTLSPolicySpec contains the TargetRefs and Validation fields. The Validation field is a BackendTLSPolicyValidation and contains CACertificateRefs, WellKnownCACertificates, and Hostname. The names of the fields were chosen to facilitate discussion, but may be substituted without blocking acceptance of the content of the API change. In fact, the CertRefs field name was changed to CACertRefs and then to CACertificateRefs as of April 2024.

The CACertificateRefs and WellKnownCACertificates fields are both optional, but one of them must be set for a valid TLS configuration. CACertificateRefs is an implementation-specific slice of named object references, each containing a single cert. We originally proposed to follow the convention established by the CertificateRefs field on Gateway , but the CertificateRef requires both a tls.key and tls.crt and a certificate reference only requires the tls.crt. WellKnownCACertificates is an optional enum that allows users to specify whether to use the set of CA certificates trusted by the Gateway (WellKnownCACertificates specified as "System"), or to use the existing CACertificateRefs (WellKnownCACertificates specified as ""). The use and definition of system certificates is implementation-dependent, and the intent is that these certificates are obtained from the underlying operating system. CACertificateRefs contains one or more references to Kubernetes objects that contain PEM-encoded TLS certificates, which are used to establish a TLS handshake between the gateway and backend pod. References to a resource in a different namespace are invalid. If ClientCertifcateRefs is unspecified, then WellKnownCACertificates must be set to "System" for a valid configuration. If WellKnownCACertificates is unspecified, then CACertificateRefs must be specified with at least one entry for a valid configuration. If WellKnownCACertficates is set to "System" and there are no system trusted certificates or the implementation doesn't define system trusted certificates, then the associated TLS connection must fail.

The Hostname field is required and is to be used to configure the SNI the Gateway should use to connect to the backend. Implementations must validate that at least one name in the certificate served by the backend matches this field. We originally proposed using a list of allowed Subject Alternative Names, but determined that this was not needed in the first round, but may be added in the future.

We originally proposed allowing the configuration of expected TLS versions, but determined that this was not needed in the first round.

Thus, the following additions would be made to the Gateway API:

//TODO: Will update this section once API changes from PR 2955 are approved.

How a client behaves

This table describes the effect that a BackendTLSPolicy has on a Route. There are only two cases where the BackendTLSPolicy will signal a Route to connect to a backend using TLS, an HTTPRoute with a backend that is targeted by a BackendTLSPolicy, either with or without listener TLS configured. (There are a few other cases where it may be possible, but is implementation dependent.)

Every implementation that claims supports for BackendTLSPolicy should document for which Routes it is being implemented.

Route Type	Gateway Config	Backend is targeted by a BackendTLSPolicy?	Connect to backend with TLS?
HTTPRoute	Listener tls	Yes	Yes
HTTPRoute	No listener tls	Yes	Yes
HTTPRoute	Listener tls	No	No
HTTPRoute	No listener tls	No	No
TLSRoute	Listener Mode: Passthrough	Yes	No
TLSRoute	Listener Mode: Terminate	Yes	Implementation-dependent
TLSRoute	Listener Mode: Passthrough	No	No
TLSRoute	Listener Mode: Terminate	No	No
TCPRoute	Listener TLS	Yes	Implementation-dependent
TCPRoute	No listener TLS	Yes	Implementation-dependent
TCPRoute	Listener TLS	No	No
TCPRoute	No listener TLS	No	No
UDPRoute	Listener TLS	Yes	No
UDPRoute	No listener TLS	Yes	No
UDPRoute	Listener TLS	No	No
UDPRoute	No listener TLS	No	No
GRPCRoute	Listener TLS	Yes	Implementation-dependent
GRPCRoute	No Listener TLS	Yes	Implementation-dependent
GRPCRoute	Listener TLS	No	No
GRPCRoute	No Listener TLS	No	No

Request Flow

Step 6 would be changed in the typical client/gateway API request flow for a gateway implemented using a reverse proxy. This is shown as bolded additions in step 6 below.

1. A client makes a request to http://foo.example.com.
2. DNS resolves the name to a Gateway address.
3. The reverse proxy receives the request on a Listener and uses the Host header to match an HTTPRoute.
4. Optionally, the reverse proxy can perform request header and/or path matching based on match rules of the HTTPRoute.
5. Optionally, the reverse proxy can modify the request, i.e. add/remove headers, based on filter rules of the HTTPRoute.
6. Lastly, the reverse proxy optionally performs a TLS handshake and forwards the request to one or more objects, i.e. Service, in the cluster based on backendRefs rules of the HTTPRoute and the TargetRefs of the BackendTLSPolicy.

Alternatives

Most alternatives are enumerated in the section "The history of backend TLS". A couple of additional alternatives are also listed here.

1. Expand BackendRef, which is already an expansion point. At first, it seems logical that since listeners are handling the client-gateway certs, BackendRefs could handle the gateway-backend certs. However, when multiple Routes to target the same Service, there would be unnecessary copying of the BackendRef every time the Service was targeted. As well, there could be multiple bBackendRefs with multiple rules on a rRoute, each of which might need the gateway-backend cert configuration, so it is not the appropriate pattern.
2. Extend HTTPRoute to indicate TLS backend support. Extending HTTPRoute would interfere with deployed implementations too much to be a practical solution.
3. Add a new type of Route for backend TLS. This is impractical because we might want to enable backend TLS on other route types in the future, and because we might want to have both TLS listeners and backend TLS on a single route.

Prior Art

TLS from gateway to backend for ingress exists in several implementations, and was developed independently.

Istio Gateway supports this with a DestinationRule:

• A secret representing a certificate/key pair, where the certificate is valid for the route host
• Set Gateway spec.servers[].port.protocol: HTTPS, spec.servers[].tls.mode=SIMPLE, spec.servers[].tls.credentialName
• Set DestinationRule spec.trafficPolicy.tls.mode: SIMPLE

Ref: Istio / Understanding TLS Configuration and Istio / Destination Rule

OpenShift Route (comparable to GW API Gateway) supports this with the following route configuration items:

• A certificate/key pair, where the certificate is valid for the route host
• A separate destination CA certificate enables the Ingress Controller to trust the destination’s certificate
• An optional, separate CA certificate that completes the certificate chain

Ref: Secured routes - Configuring Routes | Networking | OpenShift Container Platform 4.12

Contour supports this from Envoy to the backend using:

• An Envoy client certificate
• A CA certificate and SubjectName which are both used to verify the backend endpoint’s identity
• Kubernetes Service annotation: projectcontour.io/upstream-protocol.tls

Ref: Upstream TLS

GKE supports a way to encrypt traffic to the backend pods using:

• AppProtocol on Service set to HTTPS
• Load balancer does not verify the certificate used by backend pods

Ref: Secure a Gateway

Emissary supports encrypted traffic to services

• In the Mapping definition, set https:// in the spec.service field
• A spec.tls in the Mapping definition, with the name of a TLSContext
• A TLSContext to provide a client certificate, set minimum TLS version support, SNI

Ref: TLS Origination

NGINX implementation through CRDs (Comparable to Route or Policy of Gateway API) supports both TLS and mTLS

• In the Upstream section of a VirtualServer or VirtualServerRoute (equivalent to HTTPRoute) there is a simple toggle to enable TLS. This does not validate the certificate of the backend and implicitly trusts the backend in order to form the SSL tunnel. This is not about validating the certificate but obfuscating the traffic with TLS/SSL.
• A Policy attachment can be provided when certification validation is required that is called egressMTLS (egress from the proxy to the upstream). This can be tuned to perform various certificate validation tests. It was created as a Policy becuase it implies some type of AuthN/AuthZ due to the additional checks. This was also compatible with Open Service Mesh and NGINX Service Mesh and removed the need for a sidecar at the ingress controller.
• A corresponding 'IngressMTLS' policy also exists for mTLS verification of client connections to the proxy. The Policy object is used for anything that implies AuthN/AuthZ.

Ref: Upstream.TLS

Ref: EgressMTLS

Ref: IngressMTLS

Answered Questions

Q. Bowei recommended that we mention the approach of cross-namespace referencing between Route and Service. Be explicit about using the standard rules with respect to attaching policies to resources.

A. This is mentioned in the API section.

Q. Costin recommended that Gateway SHOULD authenticate with either a JWT with audience or client cert or some other means - so gateway added headers can be trusted, amongst other things.

A. This is out of scope for this proposal, which centers around application developer persona resources such as HTTPRoute and Service.

Q. Costin mentioned we need to answer the question - is configuring the connection to a backend and TLS something the route author decides - or the backend owner?

A. This is decided by the application developer persona, which would more likely, but not exclusively, be the backend owner.

Q.Costin continued, same for SAN (Subject Alternative Name) certificates. The backend owner is the application developer, and the route owner will have to collaborate with the application developer to provide the appropriate configuration for TLS. The implementation would need to take the certificate provided by the application and verify that it satisfies the requirements of the route-as-client, including SAN information. Sometimes the backend owner and route owner are the same entity.

A. This was most recently addressed by adding hostname for SNI and removing allowed SANs.

Graduation Criteria

This section is to record issues that were requested for discussion in the API section before this GEP graduates out of Provisional status.

1. Rob Scott is interested in extending the TargetRef to optionally include port, since we are targeting the entirety of a Service. See the discussion in https://github.com/kubernetes-sigs/gateway-api/pull/2113/files#r1231594914, and follow up issue in https://github.com/kubernetes-sigs/gateway-api/issues/2147
2. Michael Pleshakov asked about conflicts that could arise when multiple implementations are running in a cluster. This is a gap in our policy attachment model that needs to be addressed. See the discussion in https://github.com/kubernetes-sigs/gateway-api/pull/2113/files#r1235750540. Graduating this GEP to implementable requires an update to the Policy GEP to define how status can be nested to support multiple implementations. This will likely look very similar to Route status. See comment.
3. Rob Scott wanted to note that when this graduates to the standard channel, implementations of HTTPRoute may also be required to watch the BackendTLSPolicy. If one of these policies is attached to a Service targeted by an HTTPRoute, the implementation would be required to fully implement the policy or mark the backend invalid.

References

Gateway API TLS Use Cases

GEP-713: Metaresources and PolicyAttachment

Policy Attachment

Gateway API TLS

SIG-NET Gateway API: TLS to the K8s.Service/Backend

SAN vs SNI