Google is committed to advancing racial equity for Black communities. See how.

RFC-0107: Dynamic Offers

RFC-0107: Dynamic Offers
StatusAccepted
Areas
  • Component Framework
Description

Allow creation of offers at runtime in CFv2.

Issues
Gerrit change
Authors
Reviewers
Date submitted (year-month-day)2021-05-24
Date reviewed (year-month-day)2021-06-24

Summary

In Component Framework v2 (CFv2), a parent component can add dynamic child components to the topology at runtime using collections. However, the set of capabilities offered to such a dynamic child component cannot be configured at runtime. Offers are declared in the parent component's manifest, which is immutable by the time the parent component has started running.

The parent component's manifest treats all the dynamic children in a given collection uniformly; it has no way of even talking about an individual component within a collection. This implies certain limitations:

  1. It is not possible to offer a capability to an individual dynamic child. Offering a capability to a collection offers the capability to all dynamic children in that collection.
  2. It is not possible to offer a capability from an individual dynamic child to another component. Offering a capability from a collection creates an aggregate FIDL service that allows connecting to any dynamic child in that collection.

This document proposes dynamic offers, which can be added to the component topology at runtime.

Stakeholders

  • Users of CFv2 that build dynamic topologies at runtime. They will be represented by the Driver Framework team.
  • Component Framework area owners, who can verify that the design is sensible and doesn't conflict with any future plans.
  • Security, who can ensure this feature does not undermine the system's security properties.

Motivation

Most CFv2 use cases involve a static topology of components working together to accomplish some task. The identities and relationships between all those components are known up front and can be hard-coded into component manifests.

In some other use cases, the developer doesn't know the exact set of components that will run, but does know how these dynamic components will relate to each other. The developer can group all the dynamic components into collections, where each component in a given collection is offered the same set of capabilities.

However, some users of CFv2 want to build complex topologies at runtime, offering a unique set of capabilities to each dynamic child, and routing capabilities between dynamic children.

Use case: Drivers-as-Components

In Driver Framework's Drivers-as-Components project, dynamically-instantiated driver components depend on capabilities provided by other dynamically-instantiated driver components. We'll call capabilities that are offered from one driver component to another inter-driver capabilities.

CFv2 does not natively support inter-driver capabilities. The Driver Framework team has worked around this by implementing capability-routing-like functionality in driver_manager itself. This workaround has some pretty rough edges and exposes confusing differences between inter-driver capabilities and "normal" capabilities:

  • Components connect to "normal" capabilities by opening a path in their incoming namespace, but connect to inter-driver capabilities via a Driver-Framework-specific exposed_dir field.
  • Once given an exposed_dir, the receiving component can access all the protocols implemented by its dependency. driver_manager doesn't attenuate the receiver's access to only the protocols it needs. Driver Framework team could fix this, but doing so would further duplicate Component Framework mechanisms.
  • Driver components will need to declare the "normal" capabilities they depend on in their manifests' use section, but should not list inter-driver capabilities in the same way. If they did, Component Manager would not be able to find an offer to satisfy the use.
    • Currently, Component Manager mostly ignores these extraneous uses, but that could cause confusion in the short term, and extraneous uses might become errors in the longer term.

Driver authors should not have to see "normal" capabilities as different from inter-driver capabilities, so the distinction should be removed as soon as possible. The longer we wait, the more drivers will have to migrate when the model changes.

Design

This proposal extends the fuchsia.sys2.Realm/CreateChild API to accept a list of OfferDecls describing additional capabilities that will be offered to the newly created child, beyond those that would normally be offered. We'll refer to OfferDecls specified this way as dynamic offers. OfferDecls specified in the normal way (that is, listed in the Component Manifest) will be called static offers.

Any kind of offer (i.e., protocol, directory, etc) can be used as a dynamic offer.

Dynamic offers always target the newly created child component. As a result, the set of dynamic offers targeting a component is known at component creation time and is immutable thereafter. (Whether the target field of the OfferDecl must actually be specified or not will be determined in API review.)

Any source that would be valid for a static offer is also valid for a dynamic offer. Additionally, unlike static offers, dynamic offers can use a "sibling" dynamic child component (i.e. a component with the same parent as the target) as a source. If a dynamic offer has a dynamic child as a source, there's no restriction that the target needs to be in the same collection as the source.

Once the child component is created, the dynamic offers should behave exactly like static offers, including for considerations like component shutdown order.

Destruction

A dynamic offer will be destroyed when either of its source or its target is destroyed. After the offer has been destroyed, the system will behave exactly as if it had never existed.

Where possible, users of dynamic offers should prefer destroying targets before sources. The process of removing an offer from the topology is trivial if it's the target that's being destroyed (if an offer falls in a forest and there's no component around to hear it...). However, sometimes it can't be helped, and the source will be destroyed before the target. Notably, when using a single-run collection (introduced in RFC-0101), dynamic components are destroyed as soon as they terminate, which can happen at any time.

Even if it's the source being destroyed, the process is straightforward:

  1. Component manager shuts down the source component. This prevents the target from opening any new connections to the source, and instructs the runner to stop the source component (if it was running).
  2. When the source component stops, any channels between the source and target will close.
  3. The source component and the dynamic offer will be removed from the topology simultaneously.

From the target's perspective, the channel corresponding to the capability closes and any attempts to re-acquire the capability fail permanently, as if the capability hadn't been routed in the first place.

This RFC provides no mechanism for re-creating such an offer.

Dependency cycle prevention

CFv2 requires that components and the strong offers between them form a directed acyclic graph so that components can be gracefully shut down in dependency order. For static offers, we can ensure there are no dependency cycles by looking at the component manifests.

For dynamic offers, the following precepts are sufficient to ensure an acyclic dependency graph:

  1. Dynamic component creation order is a strict total order. In order words, no two dynamic components are added to the topology at exactly the same time.
  2. Dynamic offers are always created at the same time as their target component.
  3. All offers (both static and dynamic) are immutable.
  4. All offers (both static and dynamic) must always have a valid source and a valid target. This property is checked by CreateChild on dynamic offer creation, and whenever the source or the target is destroyed, the offer is destroyed as well.

Put together, this means that a dynamic offer's source is always older than its target. If the source was younger, that would mean it didn't exist when the target was created, and therefore the offer would have had an invalid source at creation time. Since "creation order" is a strict total order, no dependency cycles are possible.

Handling invalid routes

As of this writing, Component Manager does not proactively reject component manifests with invalid routing configuration. That is, if a capability is offered from a source that doesn't expose it or if a component doesn't receive an offer for a capability it uses, no error is raised at component creation time. An error only gets raised at the point when a component tries to open an invalidly routed capability.

In the spirit of "dynamic offers should behave like static offers", we'll implement the same behavior here. A component may create a dynamic offer for a capability that the source doesn't expose. Likewise, a parent component may create a dynamic child component without providing an offer for every capability it uses.

In the future, if Component Manager changes to return errors earlier in these situations, we will alter the contract of CreateChild to return errors as well.

Implementation

The user-facing surface of this change consists of:

  1. A backwards-compatible change to the CreateChild API to accept a list of dynamic offers as an argument.
  2. A per-collection opt-in setting to allow dynamic offers to target the components in that collection. (See the discussion in Security considerations.)
  3. An allowlist for parent components that may opt-in.

Implementation of (1) will be entirely local to Component Manager. Implementation of (2) and (3) may involve changes to tools which parse CML files. Either way, cross-team coordination should not be necessary.

Performance

This change should have no impact on system performance outside of calls to CreateChild and DestroyChild, which are rare and non-performance-critical operations. Even for these operations, no adverse performance impact is expected.

Ergonomics

The offer destruction behavior is new, and could create ergonomic challenges for authors of components whose dependencies suddenly vanish.

From the target component's perspective, determining what happened after the offer is destroyed isn't completely straightforward: it simply sees the channel close. The target component cannot immediately tell whether the source component was destroyed, stopped, or closed the channel intentionally. To determine the source was in fact destroyed, the target must attempt to re-connect and await a response from Component Manager indicating a permanent failure to route.

Security considerations

On its own, this design undermines an important security goal of CFv2: all capability routes (or potential capability routes) should be declared, so they can be audited. Any usage of dynamic offers outside of tests will need to ensure that potential capability routes can be audited via some other mechanism besides component manifests.

By using this feature, Driver Framework accepts responsibility for ensuring that the space of potential capability routes can be audited. Currently, the rules are encoded in a number of places: decentralized bind rules, source code for board drivers, ACPI, and maybe others. This makes auditing a challenge. This gap is not new and will need to be addressed by Driver Framework, but how to do so is beyond the scope of this RFC.

An opt-in setting will be added to CollectionDecl, which when present will allow the parent component to pass dynamic offers to components created in that collection. This will make it clear to a CML reviewer that dynamic offers are in play, and that sources outside the component manifest need to be consulted in order to understand the scope of potential capability routes.

Furthermore, the ability to opt-in will be allowlisted by Component Framework team, so they can ensure all usage is appropriate.

Future work

If we find more use cases for dynamic offers, we may want to introduce offer upper-bounds to the component manifest schema. Offer upper-bounds can be used to constrain the kinds of dynamic offers that can be created at runtime.

The manifest would include statements like:

  • "Protocol fuchsia.example.Foo can be dynamically offered from component A to any component in collection B", or
  • "Any directory capability can be dynamically offered between any two components in collection C".

For the initial Driver Framework use case, we consider the coarser collection-level opt-in to be sufficient. This stance MUST be re-evaluated before allowlisting any other clients.

Privacy considerations

This proposal does not expose any part of the system to data to which it didn't already have access, so it should have no impact on privacy.

Testing

Unit and integration test coverage will be added, similar to any other Component Framework feature. Additional coverage will be provided indirectly by Driver Framework's integration tests.

This feature should neither help nor harm the testability of the Component Framework API. Using collections and the Realm protocol already requires working with a real Component Manager instance in an integration testing setting, and that will continue to be the case.

Documentation

Since dynamic offers will be allowlisted and initially used only by a single client, high-level Component Framework concept documentation should not be updated to mention the existence of this feature. Documentation on the changes to CreateChild, DestroyChild, and CollectionDecl should be sufficient.

This stance MUST be re-evaluated before allowlisting any other clients.

Drawbacks, alternatives, and unknowns

The primary drawback of this design is that it makes auditing capability routes more complex, since not all potential capability routes appear in component manifests. However, this is intrinsic to Driver Framework: the topology of driver components is inherently dynamic, as devices can be added and removed from a machine on the fly. Static offers are clearly insufficient, so the story must become more complicated. Dynamic offers attempts to keep this additional complexity to a minimum by mirroring the behavior of static offers as much as possible.

Alternative: Capability tokens

Capability tokens have previously been proposed to solve a related problem. That proposal can be thought of as a generalization of this RFC, where dynamic offers can cross between realms. The parent of a source component "mints" a capability token, which it passes to another parent component (potentially via a chain of intermediate components). The receiving component can pass this token into CreateChild, which acts as an authorization mechanism to demonstrate both parents consent to the offer.

Once created, this dynamic offer should behave like a static offer, with the major exception that these dynamic offers can cross between realms. Static offers (and in the main proposal, dynamic offers as well) are always local to a single parent component.

The implications of cross-realm offers aren't fully clear, and we do not believe them to be necessary, so we have rejected this option.

Alternative: Offering "live capability providers"

fxrev.dev/483166

In the main proposal, dynamic offers are inert data. Offers represent instructions for how Component Manager ought to route a request for a capability, but don't provide access to the capability on their own.

We could instead pass around capability providers, channels that can be used to obtain capabilities directly. CreateChild would then accept a vector of named capability providers. When the dynamic child component tries to open a dynamically-offered capability, that Open request will be forwarded to the capability provider.

A parent component could obtain such a capability provider by:

  • implementing the capability provider on its own,
  • calling a Realm API to get a capability provider for one of its parent's or children's exposed capabilities, or
  • receiving a capability provider from another component over a channel.

This would unlock the same use cases as capability tokens, but has the added benefit (or debatably, downside) that parents can offer capability providers directly to their children: simply start up a server in memory and pass a capability provider to CreateChild.

However, capability providers are incompatible with many valuable properties of CFv2's routing framework: auditability of capability routes, persistence between restarts, etc.

Alternatives: Destruction

When a dynamic component that is the source of an offer is destroyed, there are a number of reasonable behaviors other than destroying the offer. This section discusses these alternatives.

Forbidding source destruction

An earlier draft of this RFC proposed simply forbidding destruction of a component if it was the source of a dynamic offer. Components would be "pinned" in this way, until the targets of those offers had been destroyed.

Even from the start, this seemed brittle and presented potential ergonomic issues (i.e. a "component could not be destroyed" error may be very difficult to handle). However, it became truly unworkable with the acceptance of RFC-0101, which introduced scenarios where components are destroyed automatically.

Destruction propagation

Another proposal was for destruction to propagate along offers. That is, if the source of a dynamic offer was destroyed, the target would be destroyed as well, and so on, recursively.

This option seemed scary and inconsistent with other CFv2 behavior. There are many circumstances where capabilities are not routable or become unavailable, and none of those circumstances cause compulsory termination or destruction of components that depend on them, even when the dependency is strong. Components can observe the fact that their dependencies are unavailable and quit if they so choose.