RFC-0171: Improved diagnostics routing

RFC-0171: Improved diagnostics routing
StatusAccepted
Areas
  • Diagnostics
  • Component Framework
Description

Introduction of CML and CMC utilities to improve diagnostics routing

Issues
Gerrit change
Authors
Reviewers
Date submitted (year-month-day)2022-05-19
Date reviewed (year-month-day)2022-06-23

Summary

A proposal to provide utilities in cmc and CML to simplify the routing of diagnostics protocols (fuchsia.logger.LogSink and fuchsia.diagnostics.InspectSink) everywhere in the tree and reduce the DX pain point of missing logs or Inspect data. While this document focuses on Inspect and Logging, it can also serve to improve availability of other protocols that most components might desire to have available, like the fuchsia.tracing.provider.Registry.

Motivation

A DX pain point of using logs is that components need to route fuchsia.logger.LogSink everywhere: in production components, tests, in RealmBuilder routes, etc. Logs are a core part of most Fuchsia experiences that we expect pretty much every component and test to use.

RFC-0168 proposes using a protocol fuchsia.inspect.InspectSink which would allow components to publish Inspect bringing some improvements and reducing technical debt. Just like fuchsia.logger.LogSink we expect all (or at least most) components to use Inspect instrumentation. Today every component can expose /diagnostics to framework which allows every component to expose Inspect and make it available to the Archivist. By moving towards a protocol, we must ensure that all components continue to be able to expose Inspect data which is highly valuable to developers to debug what their components are doing at runtime.

This is non-ergonomic and error prone as we need to update all CMLs to route this protocol manually to all components which are currently writing Inspect. LogSink has the same problem and particularly in tests it’s very easy to forget to route LogSink to the components under test causing missing logs and wasted developer time.

Component Manager leverages a component's LogSink to print routing errors attributed to that component. This improves DX as the developer can quickly spot routing errors. However, if LogSink isn't properly routed these errors end up in the global syslog attributed to Component Manager and are easier to miss by the developer looking at its own component logs.

This document attempts to improve the situation by introducing utilities in cmc and CML to simplify routing these two protocols to every component.

The Component Framework has plans to review the routing APIs and conceptualize ways to make the model more consistent and user-friendly. This will take several quarters, so an incremental approach using existing primitives and extending them is preferred.

Stakeholders

Facilitator: leannogasawara@google.com

Reviewers:

  • crjohns@google.com
  • geb@google.com
  • hjfreyer@google.com
  • shayba@google.com
  • zarvox@google.com

Consulted:

  • bryanhenry@google.com
  • dgonyeo@google.com
  • jmatt@google.com
  • thatguy@google.com

Socialization: this design was socialized in the form of a Google Docs document, a decision document where the listed alternatives were heavily discussed, and meetings and conversations among the stakeholders.

Design

To make it easier for developers to never miss logs and Inspect under the current component capability routing system, align with the principles of least privilege and hierarchical isolation, and continue to be able to expose Inspect, we will develop the following:

  • cmc required offers checker.
  • Ability to offer a capability to all children and collections in CML.
  • New CML shards for diagnostics.
  • RealmBuilder automatically offers diagnostics capabilities to all components.

cmc required offers and uses

cmc will get a command line option --must-offer-protocol containing a list of protocol names for which it will validate that the following statement is true:

For each child and collection declared in a manifest, there exists an OfferDecl from some source to it for every protocol defined in the list of required protocols.

Additionally, cmc will get an equivalent command line option --must-use-protocol that will check the equivalent, but for UseDecls.

The GN and Bazel tooling will be updated to pass fuchsia.logger.LogSink and fuchsia.diagnostics.InspectSink in these options whenever cmc is called.

Developers who wish to entirely disable this check, regardless of how cmc is called for their manifest can add the following to their CML file (this is new syntax that is introduced in CML):

{
    disable: {
        must_use_protocol: [ "fuchsia.logger.LogSink", "fuchsia.diagnostics.InspectSink" ],
        must_offer_protocol: [ "fuchsia.logger.LogSink", "fuchsia.diagnostics.InspectSink" ],
    }
}

Developers who wish to not route LogSink or InspectSink to some of their children are free to either:

  • Use optional capability routes: Route the protocol from: "void" to the child/collection for which they want to turn a single offer off.
  • Manually route the protocol from the source they desire.

The bootstrap and root realms where these capabilities originate, will need some special handling:

  • bootstrap: will have this option turned on to ensure that LogSink is routed to all components in bootstrap. Additionally, it will add an offer Inspect/LogSink from void to Archivist.
  • root: will have this option turned on to ensure we route LogSink from Archivist to all of its siblings. Since bootstrap is the one exposing this capability, we’ll add an offer Inspect/LogSink from void to bootstrap.

By doing this we hope that every developer on Fuchsia is less likely to mistakenly miss logs or Inspect data.

zarvox@ built a prototype (and relation chain) for this section.

Allow offering a capability to all children and collections in CML

To improve the DX of routing to all children we’ll introduce syntax sugar in CML to allow routing a capability to “all children and collections”.

This syntax sugar can be used as follows:

offer: [
    {
        protocol: "fuchsia.logger.LogSink",
        from: "parent",
        to: "all",
    }
]

When a CML file containing that syntax is compiled, N OfferDecls will be generated where N is the total number of collections and children that the component has.

An OfferDecl with target all will be gated in cmc for usage only for the protocols that were defined in the new optional argument described in the previous section.

CML shards

The following shards will be created:

// syslog/use.shard.cml
{
    use: [
        { protocol: "fuchsia.logger.LogSink" },
    ],
}

// syslog/offer.shard.cml
offer: [
    {
        protocol: "fuchsia.logger.LogSink",
        from: "parent",
        to: "all"
    }
]

// inspect/use.shard.cml
{
    use: [
        { protocol: "fuchsia.diagnostics.InspectSink" },
    ],
}

// inspect/offer.shard.cml
offer: [
    {
        protocol: "fuchsia.diagnostics.InspectSink",
        from: "parent",
        to: "all"
    }
]

The following existing shards will be updated:

  • syslog/client.shard.cml: includes syslog/use.shard.cml and syslog/offer.shard.cml.
  • inspect/client.shard.cml: includes inspect/use.shard.cml and inspect/offer.shard.cml.

Logical components that just perform routing and do not execute any program, can use the offer.shard.cml. Components that need to use these protocols, but need to configure what to route to their children can use the use.shard.cml. The rest can use the standard and convenient client.shard.cml.

If a component has no children or collections, but still uses the client.shard.cml (since it is using a protocol), then the offer to all statement in the shard will be a no-op given that it is syntax sugar that just expands to OfferDecls as mentioned earlier.

For convenience, we will provide a diagnostics/client.shard.cml that includes both client.shard.cml files.

RealmBuilder updates to support offer to all

To facilitate routing diagnostics protocols to all components under test, RealmBuilder will receive couple updates to allow routing a protocol to all children and collections:

  • Automatically offering LogSink and InspectSink to all children and collections. In Rust, it could look as follows:

    builder
        .add_route(
            Route::new()
                .capability(Capability::protocol_by_name("fuchsia.diagnostics.InspectSink"))
                .from(Ref::parent())
                .to(Ref::all()),
        )
        .await?;
    
  • Since we expect all tests will want to do this, except on some niche scenarios, RealmBuilder will automatically route these protocols to all components. This might seem unaligned with the approach taken in cmc and CML, but the RealmBuilder API already deviates in some areas to provide more convenient workflows that are better adapted for tests. Given that we expect that 99% of the time we will be routing these protocols to tests components, then we will teach RealmBuilder to do it automatically and provide a way to turn it off:

    let builder = RealmBuilder::new().await?;
    …
    let instance = builder
        .route_logs_to_all(false)     // defaults to true
        .route_inspect_to_all(false)
        .build()
        .await;
    

Implementation

  1. Update cmc to support the new flags and offer to all in CML.
  2. Add a syslog/offer.shard.cml that contains offer LogSink to all.
  3. Update cmc usage in tree to use the new flag and update existing CMLs that might be missing routes. The GN and Bazel SDKs will be updated, but will default to [] for the set of required protocols until OOT CMLs have been migrated to have a complete set of offers.
  4. Update cmc usage out of tree to use the new flag and update existing CMLs that might be missing routes (leveraging the offer shard).
  5. Update GN and Bazel SDKs to require the diagnostics protocols.
  6. Include the syslog/offer.shard.cml from syslog/client.shard.cml.
  7. Once rolled, refactor OOT manifests that are using the offer shard but don't need it anymore, due to it being included through the client shard.

Performance

cmc will perform some additional work, but it is not expected to have any significant impact in compile time.

Security considerations

This change aligns with Component Framework security properties, in particular the principle of least privilege and the principle of hierarchical isolation.

Privacy considerations

No privacy impact.

Testing

New cmc features will be unit tested.

Documentation

The cmc will be updated to include the new option and the CML will be updated to describe the new offer to all feature.

Drawbacks, alternatives, and unknowns

Using debug_capabilities in Environments

This was the main alternative considered. Under this alternative we would extend fuchsia.sys2.Environment to have a diagnostics_capabilities just like debug_capabilities or turn debug_capabilities into diagnostics_capabilities or just use debug_capabilities for diagnostics protocols, making them available for usage by any component in the tree from: diagnostics or from: debug.

This feature would be gated in Component Manager security policy to ensure it is only used by the root Archivist and by the Archivist embedded in tests.

Pros:

  • Every component can use InspectSink and LogSink from anywhere in the tree.
  • Aligned with the current state of the world in which every component can expose Inspect.
  • Improved DX as developers won’t need to spend time figuring out why their component isn’t logging to find out they were missing an offer in tests.
  • Possible to statically check point to point usage of the capability.
  • All components start with “nothing” in their namespace unless they explicitly use the capability.
  • Covers dynamic components created through fuchsia.component.Realm/CreateChild.

Cons:

  • No more explicit parent-child offering, meaning, this is not aligned with the security principle of hierarchical isolation.
  • Replacing or mocking LogSink/InspectSink in the topology requires adjusting the environment which requires a security policy change.
  • Isn’t eating our own dogfood – third-party developers don’t get to use environments for arbitrary purposes for their protocols, why do we?

Make LogSink and InspectSink framework capabilities

Allow these protocols to be used from: framework. The Archivist can either expose these capabilities to to framework or there can be a contract between Archivist and Component Manager to serve these capabilities.

Pros:

  • Same pros as in the previous alternative with the detail that diagnostics protocols (InspectSink and LogSink) become framework protocols even if not served by Component Manager.
  • Removes the need for capability attribution, as attribution would happen directly in the framework since each component gets its own unique set of framework capabilities.

Cons:

  • Same cons as the previous alternative.
  • First instance of a protocol used from framework that isn’t offered directly by Component Manager : itself.
  • Not clear how to provide isolated logs in tests without building an additional mechanism to be used by Test Manager.
  • Establishes a single log destination for all components, device-wide.

Have cmc offer LogSink to all children automatically

Instead of having a flag that requests the user to add the OfferDecl to their CML, cmc would do this automatically for every child and collection.

Pros:

  • Explicit parent-child offering which can help with mocking, replacing the protocol in the topology, etc.
  • No changes in CML.

Cons:

  • Special treatment of a capability in CMC.
  • Inconsistency in behavior of components declared in .cml and components constructed via Realm/CreateChild.

Providing an option in cmc and syntax sugar in CML makes this more flexible and provides a mechanism that others could take advantage of, not only for diagnostics.

Do nothing and route as usual

Have developers manually offer LogSink and InspectSink to all their children.

Pros:

  • Explicit parent-child offering which can help with mocking, replacing the protocol in the topology, etc.
  • API boundaries remain a local concern between parent and child, rather than one that involves other parties.

Cons:

  • Current problems: it’s easy to miss routing LogSink leading to lost time while debugging tests.
  • Additional problems: easy to miss routing InspectSink to some component (given that today everyone can expose it) leading to missing diagnostics in the field. This is just the same problem LogSink has, so now we have it in two protocols rather than only one.

Given how widely used these protocols are, we believe that adding additional options in cmc and CML help reduce the likelihood of missing a route.

Other ideas

Other ideas were discussed such as using capability bundles to route a diagnostics bundle containing both protocols or improving environments in the form of domains or capability sources. These were discarded in favor of a near term solution using existing mechanisms and APIs given the plans to review routing APIs.

Prior art and references

N/A