Dynamically available HALs

Implementing dynamic shutdown involves wiring up data flows and executing dynamic processes as detailed in the following sections.

Changes to HAL definitions

Dynamic shutdown requires information on which processes serve what HAL interfaces (this information may also be useful later in other contexts) as well as not starting processes on boot and not restarting them (until requested again) when they exit.

# some init.rc script associated with the HAL
service vendor.some-service-name /vendor/bin/hw/some-binary-service
    # init language extension, provides information of what service is served
    # if multiple interfaces are served, they can be specified one on each line
    interface android.hardware.light@2.0::ILight default
    # restarted if hwservicemanager dies
    # would also cause the hal to start early during boot if disabled wasn't set
    class hal
    # will not be restarted if it exits until it is requested to be restarted
    oneshot
    # will only be started when requested
    disabled
    # ... other properties

Changes to init and hwservicemanager

Dynamic shutdown also requires the hwservicemanager to tell init to start requested services. In Android 9, init includes three additional control messages (e.g. ctl.start): ctl.interface_start, ctl.interface_stop, and ctl.interface_restart. These messages can be used to signal init to bring up and down specific hardware interfaces. When a service is requested and isn't registered, hwservicemanager requests that the service be started. However, dynamic HALs don't require using any of these.

Determine HAL exit

In Android 9, HAL exit has to be manually determined. For Android 10 and higher, it can also be determined with automatic lifecycles.

Dynamic shutdown requires multiple policies for deciding when to start a HAL and when to shutdown a HAL. If a HAL decides to exit for any reason, it will automatically be restarted when it is needed again using the information provided in the HAL definition and the infrastructure provided by changes to init and hwservicemanager. This could involve a couple of different strategies, including:

• A HAL could choose to call exit on itself if someone calls a close or similar API on it. This behavior must be specified in the corresponding HAL interface.
• HALs can shut down when their task is completed (documented in the HAL file).

Automatic lifecycles

Android 10 adds more support to the kernel and hwservicemanager, which allows HALs to shut down automatically whenever they have no clients. To use this feature, do all of the steps in Changes to HAL definitions as well as:

• Register the service in C++ with LazyServiceRegistrar instead of the member function, registerAsService, for example:

  // only one instance of LazyServiceRegistrar per process
  LazyServiceRegistrar registrar;
  registrar.registerAsService(myHidlService /* , "default" */);

• Verify that the HAL client keeps a reference to the top-level HAL (the interface registered with hwservicemanager) only when it's in use. To avoid delays if this reference is dropped on a hwbinder thread that continues to execute, the client should also call IPCThreadState::self()->flushCommands() after dropping the reference to ensure that the binder driver is notified of the associated reference count changes.