Policy Enforcement and Authorization

As most OpenStack projects, Neutron leverages oslo_policy [1]. However, since Neutron loves to be special and complicate every developer’s life, it also “augments” oslo_policy capabilities by:

  • A wrapper module with its own API: neutron.policy;

  • The ability of adding fine-grained checks on attributes for resources in request bodies;

  • The ability of using the policy engine to filter out attributes in responses;

  • Adding some custom rule checks beyond those defined in oslo_policy;

This document discusses Neutron-specific aspects of policy enforcement, and in particular how the enforcement logic is wired into API processing. For any other information please refer to the developer documentation for oslo_policy [2].

Authorization workflow

The Neutron API controllers perform policy checks in two phases during the processing of an API request:

  • Request authorization, immediately before dispatching the request to the plugin layer for POST, PUT, and DELETE, and immediately after returning from the plugin layer for GET requests;

  • Response filtering, when building the response to be returned to the API consumer.

Request authorization

The aim of this step is to authorize processing for a request or reject it with an error status code. This step uses the neutron.policy.enforce routine. This routine raises oslo_policy.PolicyNotAuthorized when policy enforcement fails. The Neutron REST API controllers catch this exception and return:

  • A 403 response code on a POST request or an PUT request for an object owned by the project submitting the request;

  • A 403 response for failures while authorizing API actions such as add_router_interface;

  • A 404 response for DELETE, GET and all other PUT requests.

For DELETE operations the resource must first be fetched. This is done invoking the same _item [3] method used for processing GET requests. This is also true for PUT operations, since the Neutron API implements PATCH semantics for PUTs. The criteria to evaluate are built in the _build_match_rule [4] routine. This routine takes in input the following parameters:

  • The action to be performed, in the <operation>_<resource> form, e.g.: create_network

  • The data to use for performing checks. For POST operations this could be a partial specification of the object, whereas it is always a full specification for GET, PUT, and DELETE requests, as resource data are retrieved before dispatching the call to the plugin layer.

  • The collection name for the resource specified in the previous parameter; for instance, for a network it would be the “networks”.

The _build_match_rule routine returns a oslo_policy.RuleCheck instance built in the following way:

  • Always add a check for the action being performed. This will match a policy like create_network in policy.yaml;

  • Return for GET operations; more detailed checks will be performed anyway when building the response;

  • For each attribute which has been explicitly specified in the request create a rule matching policy names in the form <operation>_<resource>:<attribute> rule, and link it with the previous rule with an ‘And’ relationship (using oslo_policy.AndCheck); this step will be performed only if the enforce_policy flag is set to True in the resource attribute descriptor (usually found in a data structure called RESOURCE_ATTRIBUTE_MAP);

  • If the attribute is a composite one then further rules will be created; These will match policy names in the form <operation>_<resource>:<attribute>:<sub_attribute>. An ‘And’ relationship will be used in this case too.

As all the rules to verify are linked by ‘And’ relationships, all the policy checks should succeed in order for a request to be authorized. Rule verification is performed by oslo_policy with no “customization” from the Neutron side.

Response Filtering

Some Neutron extensions, like the provider networks one, add some attribute to resources which are however not meant to be consumed by all clients. This might be because these attributes contain implementation details, or are meant only to be used when exchanging information between services, such as Nova and Neutron;

For this reason the policy engine is invoked again when building API responses. This is achieved by the _exclude_attributes_by_policy [5] method in neutron.api.v2.base.Controller;

This method, for each attribute in the response returned by the plugin layer, first checks if the is_visible flag is True. In that case it proceeds to checking policies for the attribute; if the policy check fails the attribute is added to a list of attributes that should be removed from the response before returning it to the API client.

The neutron.policy API

The neutron.policy module exposes a simple API whose main goal if to allow the REST API controllers to implement the authorization workflow discussed inu this document. It is a bad practice to call the policy engine from within the plugin layer, as this would make request authorization dependent on configured plugins, and therefore make API behaviour dependent on the plugin itself, which defies Neutron tenet of being backend agnostic.

The neutron.policy API exposes the following routines:

  • init Initializes the policy engine loading rules from the json policy (files). This method can safely be called several times.

  • reset Clears all the rules currently configured in the policy engine. It is called in unit tests and at the end of the initialization of core API router [6] in order to ensure rules are loaded after all the extensions are loaded.

  • refresh Combines init and reset. Called when a SIGHUP signal is sent to an API worker.

  • set_rules Explicitly set policy engine’s rules. Used only in unit tests.

  • check Perform a check using the policy engine. Builds match rules as described in this document, and then evaluates the resulting rule using oslo_policy’s policy engine. Returns True if the checks succeeds, false otherwise.

  • enforce Operates like the check routine but raises if the check in oslo_policy fails.

Neutron specific policy rules

Neutron provides two additional policy rule classes in order to support the “augmented” authorization capabilities it provides. They both extend oslo_policy.RuleCheck and are registered using the oslo_policy.register decorator.

OwnerCheck: Extended Checks for Resource Ownership

This class is registered for rules matching the tenant_id keyword and overrides the generic check performed by oslo_policy in this case. It uses for those cases where neutron needs to check whether the project submitting a request for a new resource owns the parent resource of the one being created. Current usages of OwnerCheck include, for instance, creating and updating a subnet. This class supports the extension parent resources owner check which the parent resource introduced by service plugins. Such as router and floatingip owner check for router service plugin. Developers can register the extension resource name and service plugin name which were registered in neutron-lib into EXT_PARENT_RESOURCE_MAPPING which is located in neutron_lib.services.constants.

The check, performed in the __call__ method, works as follows:

  • verify if the target field is already in the target data. If yes, then simply verify whether the value for the target field in target data is equal to value for the same field in credentials, just like oslo_policy.GenericCheck would do. This is also the most frequent case as the target field is usually tenant_id;

  • if the previous check failed, extract a parent resource type and a parent field name from the target field. For instance networks:tenant_id identifies the tenant_id attribute of the network resource. For extension parent resource case, ext_parent:tenant_id identifies the tenant_id attribute of the registered extension resource in EXT_PARENT_RESOURCE_MAPPING;

  • if no parent resource or target field could be identified raise a PolicyCheckError exception;

  • Retrieve a ‘parent foreign key’ from the _RESOURCE_FOREIGN_KEYS data structure in neutron.policy. This foreign key is simply the attribute acting as a primary key in the parent resource. A PolicyCheckError exception will be raised if such ‘parent foreign key’ cannot be retrieved;

  • Using the core plugin, retrieve an instance of the resource having ‘parent foreign key’ as an identifier;

  • Finally, verify whether the target field in this resource matches the one in the initial request data. For instance, for a port create request, verify whether the tenant_id of the port data structure matches the tenant_id of the network where this port is being created.

FieldCheck: Verify Resource Attributes

This class is registered with the policy engine for rules matching the ‘field’ keyword, and provides a way to perform fine grained checks on resource attributes. For instance, using this class of rules it is possible to specify a rule for granting every project read access to shared resources.

In policy.yaml, a FieldCheck rules is specified in the following way:

> field: <resource>:<field>=<value>

This will result in the initialization of a FieldCheck that will check for <field> in the target resource data, and return True if it is equal to <value> or return False is the <field> either is not equal to <value> or does not exist at all.

Guidance for Neutron API developers

When developing REST APIs for Neutron it is important to be aware of how the policy engine will authorize these requests. This is true both for APIs served by Neutron “core” and for the APIs served by the various Neutron “stadium” services.

  • If an attribute of a resource might be subject to authorization checks then the enforce_policy attribute should be set to True. While setting this flag to True for each attribute is a viable strategy, it is worth noting that this will require a call to the policy engine for each attribute, thus consistently increasing the time required to complete policy checks for a resource. This could result in a scalability issue, especially in the case of list operations retrieving a large number of resources;

  • Some resource attributes, even if not directly used in policy checks might still be required by the policy engine. This is for instance the case of the tenant_id attribute. For these attributes the required_by_policy attribute should always set to True. This will ensure that the attribute is included in the resource data sent to the policy engine for evaluation;

  • The tenant_id attribute is a fundamental one in Neutron API request authorization. The default policy, admin_or_owner, uses it to validate if a project owns the resource it is trying to operate on. To this aim, if a resource without a tenant_id is created, it is important to ensure that ad-hoc authZ policies are specified for this resource.

  • There is still only one check which is hardcoded in Neutron’s API layer: the check to verify that a project owns the network on which it is creating a port. This check is hardcoded and is always executed when creating a port, unless the network is shared. Unfortunately a solution for performing this check in an efficient way through the policy engine has not yet been found. Due to its nature, there is no way to override this check using the policy engine.

  • It is strongly advised to not perform policy checks in the plugin or in the database management classes. This might lead to divergent API behaviours across plugins. Also, it might leave the Neutron DB in an inconsistent state if a request is not authorized after it has already been dispatched to the backend.

Notes

  • No authorization checks are performed for requests coming from the RPC over AMQP channel. For all these requests a neutron admin context is built, and the plugins will process them as such.

  • For PUT and DELETE requests a 404 error is returned on request authorization failures rather than a 403, unless the project submitting the request own the resource to update or delete. This is to avoid conditions in which an API client might try and find out other projects’ resource identifiers by sending out PUT and DELETE requests for random resource identifiers.

  • There is no way at the moment to specify an OR relationship between two attributes of a given resource (eg.: port.name == 'meh' or port.status == 'DOWN'), unless the rule with the or condition is explicitly added to the policy.yaml file.

  • OwnerCheck performs a plugin access; this will likely require a database access, but since the behaviour is implementation specific it might also imply a round-trip to the backend. This class of checks, when involving retrieving attributes for ‘parent’ resources should be used very sparingly.

  • In order for OwnerCheck rules to work, parent resources should have an entry in neutron.policy._RESOURCE_FOREIGN_KEYS; moreover the resource must be managed by the ‘core’ plugin (ie: the one defined in the core_plugin configuration variable)

Policy-in-Code support

Guideline on defining in-code policies

The following is the guideline of policy definitions.

Ideally we should define all available policies, but in the neutron policy enforcement it is not practical to define all policies because we check all attributes of a target resource in the Response Filtering. Considering this, we have the special guidelines for “get” operation.

  • All policies of <action>_<resource> must be defined for all types of operations. Valid actions are create, update, delete and get.

  • get_<resourceS> (get plural) is unnecessary. The neutron API layer use a single form policy get_<resource> when listing resources [7] [8].

  • Member actions for individual resources must be defined. For example, add_router_interface of router resource.

  • All policies with attributes on “create”, “update” and “delete” actions must be defined. <action>_<resource>:<attribute>(:<sub_attribute>) policy is required for attributes with enforce_policy in the API definitions. Note that it is recommended to define even if a rule is same as for <action>_<resource> from the documentation perspective.

  • For a policy with attributes of “get” actions like get_<resource>:<attribute>(:<sub_attribute>), the following guideline is applied:

    • A policy with an attribute must be defined if the policy is different from the policy for get_<resource> (without attributes).

    • If a policy with an attribute is same as for get_<resource>, there is no need to define it explicitly. This is for simplicity. We check all attributes of a target resource in the process of Response Filtering so it leads to a long long policy definitions for “get” actions in our documentation. It is not happy for operators either.

    • If an attribute is marked as enforce_policy, it is recommended to define the corresponding policy with the attribute. This is for clarification. If an attribute is marked as enforce_policy in the API definitions, for example, the neutron API limits to set such attribute only to admin users but allows to retrieve a value for regular users. If policies for the attribute are different across the types of operations, it is better to define all of them explicitly.

References