Quota Management and Enforcement

Most resources exposed by the Neutron API are subject to quota limits. The Neutron API exposes an extension for managing such quotas. Quota limits are enforced at the API layer, before the request is dispatched to the plugin.

Default values for quota limits are specified in neutron.conf. Admin users can override those defaults values on a per-tenant basis. Limits are stored in the Neutron database; if no limit is found for a given resource and tenant, then the default value for such resource is used. Configuration-based quota management, where every tenant gets the same quota limit specified in the configuration file, has been deprecated as of the Liberty release.

Please note that Neutron does not support both specification of quota limits per user and quota management for hierarchical multitenancy (as a matter of fact Neutron does not support hierarchical multitenancy at all). Also, quota limits are currently not enforced on RPC interfaces listening on the AMQP bus.

Plugin and ML2 drivers are not supposed to enforce quotas for resources they manage. However, the subnet_allocation [1] extension is an exception and will be discussed below.

The quota management and enforcement mechanisms discussed here apply to every resource which has been registered with the Quota engine, regardless of whether such resource belongs to the core Neutron API or one of its extensions.

High Level View

There are two main components in the Neutron quota system:

  • The Quota API extension;
  • The Quota Engine.

Both components rely on a quota driver. The neutron codebase currently defines two quota drivers:

  • neutron.db.quota.driver.DbQuotaDriver
  • neutron.quota.ConfDriver

The latter driver is however deprecated.

The Quota API extension handles quota management, whereas the Quota Engine component handles quota enforcement. This API extension is loaded like any other extension. For this reason plugins must explicitly support it by including “quotas” in the support_extension_aliases attribute.

In the Quota API simple CRUD operations are used for managing tenant quotas. Please note that the current behaviour when deleting a tenant quota is to reset quota limits for that tenant to configuration defaults. The API extension does not validate the tenant identifier with the identity service.

Performing quota enforcement is the responsibility of the Quota Engine. RESTful API controllers, before sending a request to the plugin, try to obtain a reservation from the quota engine for the resources specified in the client request. If the reservation is successful, then it proceeds to dispatch the operation to the plugin.

For a reservation to be successful, the total amount of resources requested, plus the total amount of resources reserved, plus the total amount of resources already stored in the database should not exceed the tenant’s quota limit.

Finally, both quota management and enforcement rely on a “quota driver” [2], whose task is basically to perform database operations.

Quota Management

The quota management component is fairly straightforward.

However, unlike the vast majority of Neutron extensions, it uses it own controller class [3]. This class does not implement the POST operation. List, get, update, and delete operations are implemented by the usual index, show, update and delete methods. These method simply call into the quota driver for either fetching tenant quotas or updating them.

The _update_attributes method is called only once in the controller lifetime. This method dynamically updates Neutron’s resource attribute map [4] so that an attribute is added for every resource managed by the quota engine. Request authorisation is performed in this controller, and only ‘admin’ users are allowed to modify quotas for tenants. As the neutron policy engine is not used, it is not possible to configure which users should be allowed to manage quotas using policy.json.

The driver operations dealing with quota management are:

  • delete_tenant_quota, which simply removes all entries from the ‘quotas’ table for a given tenant identifier;
  • update_quota_limit, which adds or updates an entry in the ‘quotas’ tenant for a given tenant identifier and a given resource name;
  • _get_quotas, which fetches limits for a set of resource and a given tenant identifier
  • _get_all_quotas, which behaves like _get_quotas, but for all tenants.

Resource Usage Info

Neutron has two ways of tracking resource usage info:

  • CountableResource, where resource usage is calculated every time quotas limits are enforced by counting rows in the resource table and reservations for that resource.
  • TrackedResource, which instead relies on a specific table tracking usage data, and performs explicitly counting only when the data in this table are not in sync with actual used and reserved resources.

Another difference between CountableResource and TrackedResource is that the former invokes a plugin method to count resources. CountableResource should be therefore employed for plugins which do not leverage the Neutron database. The actual class that the Neutron quota engine will use is determined by the track_quota_usage variable in the quota configuration section. If True, TrackedResource instances will be created, otherwise the quota engine will use CountableResource instances. Resource creation is performed by the create_resource_instance factory method in the neutron.quota.resource module.

From a performance perspective, having a table tracking resource usage has some advantages, albeit not fundamental. Indeed the time required for executing queries to explicitly count objects will increase with the number of records in the table. On the other hand, using TrackedResource will fetch a single record, but has the drawback of having to execute an UPDATE statement once the operation is completed. Nevertheless, CountableResource instances do not simply perform a SELECT query on the relevant table for a resource, but invoke a plugin method, which might execute several statements and sometimes even interacts with the backend before returning. Resource usage tracking also becomes important for operational correctness when coupled with the concept of resource reservation, discussed in another section of this chapter.

Tracking quota usage is not as simple as updating a counter every time resources are created or deleted. Indeed a quota-limited resource in Neutron can be created in several ways. While a RESTful API request is the most common one, resources can be created by RPC handlers listing on the AMQP bus, such as those which create DHCP ports, or by plugin operations, such as those which create router ports.

To this aim, TrackedResource instances are initialised with a reference to the model class for the resource for which they track usage data. During object initialisation, SqlAlchemy event handlers are installed for this class. The event handler is executed after a record is inserted or deleted. As result usage data for that resource and will be marked as ‘dirty’ once the operation completes, so that the next time usage data is requested, it will be synchronised counting resource usage from the database. Even if this solution has some drawbacks, listed in the ‘exceptions and caveats’ section, it is more reliable than solutions such as:

  • Updating the usage counters with the new ‘correct’ value every time an operation completes.
  • Having a periodic task synchronising quota usage data with actual data in the Neutron DB.

Finally, regardless of whether CountableResource or TrackedResource is used, the quota engine always invokes its count() method to retrieve resource usage. Therefore, from the perspective of the Quota engine there is absolutely no difference between CountableResource and TrackedResource.

Quota Enforcement

Before dispatching a request to the plugin, the Neutron ‘base’ controller [5] attempts to make a reservation for requested resource(s). Reservations are made by calling the make_reservation method in neutron.quota.QuotaEngine. The process of making a reservation is fairly straightforward:

  • Get current resource usages. This is achieved by invoking the count method on every requested resource, and then retrieving the amount of reserved resources.
  • Fetch current quota limits for requested resources, by invoking the _get_tenant_quotas method.
  • Fetch expired reservations for selected resources. This amount will be subtracted from resource usage. As in most cases there won’t be any expired reservation, this approach actually requires less DB operations than doing a sum of non-expired, reserved resources for each request.
  • For each resource calculate its headroom, and verify the requested amount of resource is less than the headroom.
  • If the above is true for all resource, the reservation is saved in the DB, otherwise an OverQuotaLimit exception is raised.

The quota engine is able to make a reservation for multiple resources. However, it is worth noting that because of the current structure of the Neutron API layer, there will not be any practical case in which a reservation for multiple resources is made. For this reason performance optimisation avoiding repeating queries for every resource are not part of the current implementation.

In order to ensure correct operations, a row-level lock is acquired in the transaction which creates the reservation. The lock is acquired when reading usage data. In case of write-set certification failures, which can occur in active/active clusters such as MySQL galera, the decorator oslo_db.api.wrap_db_retry will retry the transaction if a DBDeadLock exception is raised. While non-locking approaches are possible, it has been found out that, since a non-locking algorithms increases the chances of collision, the cost of handling a DBDeadlock is still lower than the cost of retrying the operation when a collision is detected. A study in this direction was conducted for IP allocation operations, but the same principles apply here as well [6]. Nevertheless, moving away for DB-level locks is something that must happen for quota enforcement in the future.

Committing and cancelling a reservation is as simple as deleting the reservation itself. When a reservation is committed, the resources which were committed are now stored in the database, so the reservation itself should be deleted. The Neutron quota engine simply removes the record when cancelling a reservation (ie: the request failed to complete), and also marks quota usage info as dirty when the reservation is committed (ie: the request completed correctly). Reservations are committed or cancelled by respectively calling the commit_reservation and cancel_reservation methods in neutron.quota.QuotaEngine.

Reservations are not perennial. Eternal reservation would eventually exhaust tenants’ quotas because they would never be removed when an API worker crashes whilst in the middle of an operation. Reservation expiration is currently set to 120 seconds, and is not configurable, not yet at least. Expired reservations are not counted when calculating resource usage. While creating a reservation, if any expired reservation is found, all expired reservation for that tenant and resource will be removed from the database, thus avoiding build-up of expired reservations.

Setting up Resource Tracking for a Plugin

By default plugins do not leverage resource tracking. Having the plugin explicitly declare which resources should be tracked is a precise design choice aimed at limiting as much as possible the chance of introducing errors in existing plugins.

For this reason a plugin must declare which resource it intends to track. This can be achieved using the tracked_resources decorator available in the neutron.quota.resource_registry module. The decorator should ideally be applied to the plugin’s __init__ method.

The decorator accepts in input a list of keyword arguments. The name of the argument must be a resource name, and the value of the argument must be a DB model class. For example:

port=models_v2.Port, subnet=models_v2.Subnet, subnetpool=models_v2.SubnetPool)

Will ensure network, port, subnet and subnetpool resources are tracked. In theory, it is possible to use this decorator multiple times, and not exclusively to __init__ methods. However, this would eventually lead to code readability and maintainability problems, so developers are strongly encourage to apply this decorator exclusively to the plugin’s __init__ method (or any other method which is called by the plugin only once during its initialization).

Notes for Implementors of RPC Interfaces and RESTful Controllers

Neutron unfortunately does not have a layer which is called before dispatching the operation from the plugin which can be leveraged both from RESTful and RPC over AMQP APIs. In particular the RPC handlers call straight into the plugin, without doing any request authorisation or quota enforcement.

Therefore RPC handlers must explicitly indicate if they are going to call the plugin to create or delete any sort of resources. This is achieved in a simple way, by ensuring modified resources are marked as dirty after the RPC handler execution terminates. To this aim developers can use the mark_resources_dirty decorator available in the module neutron.quota.resource_registry.

The decorator would scan the whole list of registered resources, and store the dirty status for their usage trackers in the database for those resources for which items have been created or destroyed during the plugin operation.

Exceptions and Caveats

Please be aware of the following limitations of the quota enforcement engine:

  • Subnet allocation from subnet pools, in particularly shared pools, is also subject to quota limit checks. However this checks are not enforced by the quota engine, but trough a mechanism implemented in the neutron.ipam.subnetalloc module. This is because the Quota engine is not able to satisfy the requirements for quotas on subnet allocation.
  • The quota engine also provides a limit_check routine which enforces quota checks without creating reservations. This way of doing quota enforcement is extremely unreliable and superseded by the reservation mechanism. It has not been removed to ensure off-tree plugins and extensions which leverage are not broken.
  • SqlAlchemy events might not be the most reliable way for detecting changes in resource usage. Since the event mechanism monitors the data model class, it is paramount for a correct quota enforcement, that resources are always created and deleted using object relational mappings. For instance, deleting a resource with a query.delete call, will not trigger the event. SQLAlchemy events should be considered as a temporary measure adopted as Neutron lacks persistent API objects.
  • As CountableResource instance do not track usage data, when making a reservation no write-intent lock is acquired. Therefore the quota engine with CountableResource is not concurrency-safe.
  • The mechanism for specifying for which resources enable usage tracking relies on the fact that the plugin is loaded before quota-limited resources are registered. For this reason it is not possible to validate whether a resource actually exists or not when enabling tracking for it. Developers should pay particular attention into ensuring resource names are correctly specified.
  • The code assumes usage trackers are a trusted source of truth: if they report a usage counter and the dirty bit is not set, that counter is correct. If it’s dirty than surely that counter is out of sync. This is not very robust, as there might be issues upon restart when toggling the use_tracked_resources configuration variable, as stale counters might be trusted upon for making reservations. Also, the same situation might occur if a server crashes after the API operation is completed but before the reservation is committed, as the actual resource usage is changed but the corresponding usage tracker is not marked as dirty.