Convert Banjo protocols to FIDL protocols

This page provides instructions, best practices, and examples related to converting Banjo protocols to FIDL protocols as part of DFv1-to-DFv2 driver migration.

Update the DFv1 driver from Banjo to FIDL

Updating a driver's .fidl file is a good starting point for driver migration because everything originates from the .fidl file. Fortunately, Banjo and FIDL generate code from the same IDL (which is FIDL), so you may not need to make significant changes to the existing .fidl file.

The example .fidl files below show the changes before and after the Banjo-to-FIDL migration:

To update your DFv1 driver from Banjo to FIDL, make the following changes (mostly in the driver's .fidl file):

  1. Update attributes before protocol definitions.
  2. Update function definitions to use FIDL error syntax.
  3. Update FIDL targets in BUILD.gn.
  4. Move the FIDL file to the SDK/FIDL directory.

1. Update attributes before protocol definitions

To use Banjo, the @transport("Banjo") attribute is required in a .fidl file. But this attribute is not necessary for FIDL (because FIDL is the default). So you can just delete the @transport("Banjo") attribute from your driver's .fidl file, for example:

  • From using Banjo:

    @transport("Banjo")
    @banjo_layout("ddk-protocol")
    protocol MyExampleProtocol {
     ...
    }
    
  • To using FIDL:

    @discoverable
    @transport("Driver")
    protocol MyExampleProtocol {
     ...
    }
    

    In the example above, the @discoverable attribute is required for all FIDL protocols. This attribute allows the client to search this protocol using its generated name.

    However, the @transport("Driver") attribute (which indicates that this is a Driver transport protocol) is optional for drivers that use the driver runtime FIDL. And driver runtime migration is only necessary if your driver talks to other drivers co-located in the same driver host.

    For more examples on the Driver transport protocols, see this Driver transport examples directory.

2. Update function definitions to use FIDL error syntax

If some function definitions in the .fidl file include return status, you need to update them to use FIDL error syntax instead of adding status in the return structure, for example:

  • From using a return structure:

    protocol MyExampleProtocol {
       MyExampleFunction() -> (struct {
           Error zx.status;
       });
    };
    

    (Source: wlanphy-impl.fidl)

  • To returning FIDL error syntax:

    protocol MyExampleProtocol {
       BMyExampleFunction() -> () error zx.status;
    };
    

    (Source: phyimpl.fidl)

Using FIDL error syntax has the following benefits:

  • The return structure can now focus on the data that needs to be sent to the server end.

  • The error handling on the server end is cleaner because fetching the status does not require reading into the return structure.

3. Update FIDL targets in BUILD.gn

Edit the BUILD.gn file (of this .fidl file) to add the following line in the FIDL targets:

contains_drivers = true

The example below shows the contains_drivers = true line in a FIDL target:

import("//build/fidl/fidl.gni")

fidl("fuchsia.wlan.phyimpl") {
 sdk_category = "partner"
 sources = [ "phyimpl.fidl" ]
 public_deps = [
   "//sdk/fidl/fuchsia.wlan.common",
   "//sdk/fidl/fuchsia.wlan.ieee80211",
   "//zircon/vdso/zx",
 ]
 contains_drivers = true
 enable_banjo = true
}

(Source: BUILD.gn)

Without the contains_drivers = true line, the protocols with @transport("Driver") attribute won't be generated into FIDL libraries properly.

4. Move the FIDL file to the SDK/FIDL directory

Move the updated .fidl file from the sdk/banjo directory to the sdk/fidl directory.

The sdk/fidl directory is the default location for storing all files that generate FIDL code. However, if Banjo structures or functions are still used anywhere else (that is, outside of driver communication), moving this .fidl file is not allowed. In that case, you can keep a copy of this .fidl file in the sdk/banjo directory.

(Optional) Update the DFv1 driver to use the driver runtime

This section requires that the updated .fidl file from the previous section can successfully generate FIDL code. This FIDL code is used to establish the driver runtime FIDL communication between drivers.

To update a DFv1 driver to use the driver runtime, the steps are:

  1. Update dependencies for the driver runtime.
  2. Set up client and server objects for the driver runtime FIDL.
  3. Update the driver to use the driver runtime FIDL.
  4. Make FIDL requests.

1. Update dependencies for the driver runtime

Update the server and client ends to include new dependencies for using the driver runtime:

  1. In the BUILD.gn file, update the dependencies fields to include the following lines:

    //sdk/fidl/<YOUR_FIDL_LIB>_cpp_wire
    //sdk/fidl/<YOUR_FIDL_LIB>_cpp_driver
    //src/devices/lib/driver:driver_runtime
    

    Replace YOUR_FIDL_LIB with the name of your FIDL library, for example:

    public_deps = [
      ...
      "//sdk/fidl/fuchsia.factory.wlan:fuchsia.factory.wlan_cpp_wire",
      "//sdk/fidl/fuchsia.wlan.fullmac:fuchsia.wlan.fullmac_cpp_driver",
      "//sdk/fidl/fuchsia.wlan.phyimpl:fuchsia.wlan.phyimpl_cpp_driver",
      ...
      "//src/devices/lib/driver:driver_runtime",
      ...
    ]
    

    (Source: BUILD.gn)

  2. In the headers of source code, update the include lines, for example:

    #include <fidl/<YOUR_FIDL_LIB>/cpp/driver/wire.h>
    #include <lib/fdf/cpp/arena.h>
    #include <lib/fdf/cpp/channel.h>
    #include <lib/fdf/cpp/channel_read.h>
    #include <lib/fdf/cpp/dispatcher.h>
    #include <lib/fidl/llcpp/connect_service.h>
    #include <lib/fidl/llcpp/vector_view.h>
    ...
    

    (Source: wlanphy-impl-device.h)

2. Set up client and server objects for the driver runtime FIDL

Update the server and client ends to use the driver runtime FIDL.

On the client end, do the following:

  1. Declare a FIDL client object (fdf::WireSharedClient<ProtocolName> or fdf::WireSyncClient<ProtocolName>) in the device class.

    This FIDL client object enables you to make FIDL calls that send requests to the server end.

    The example code below shows a FIDL client object in a device class:

    class Device : public fidl::WireServer<fuchsia_wlan_device::Phy>,
                   public ::ddk::Device<Device, ::ddk::MessageableManual, ::ddk::Unbindable> {
     ...
     private:
      // Dispatcher for being a FIDL server listening MLME requests.
      async_dispatcher_t* server_dispatcher_;
    
      // The FIDL client to communicate with iwlwifi
      fdf::WireSharedClient<fuchsia_wlan_wlanphyimpl::WlanphyImpl> client_;
     ...
    

    (Source: device_dfv2.h)

  2. (For asynchronous calls only) Declare a dispatcher object (fdf::Dispatcher) in the device class.

    A dispatcher is needed to bind the FIDL client object from step 1.

    The example code below shows the FIDL client and dispatcher objects in the device class:

    class Device : public fidl::WireServer<fuchsia_wlan_device::Phy>,
                   public ::ddk::Device<Device, ::ddk::MessageableManual, ::ddk::Unbindable> {
     ...
     private:
      // Dispatcher for being a FIDL server listening MLME requests.
      async_dispatcher_t* server_dispatcher_;
    
      // The FIDL client to communicate with iwlwifi
      fdf::WireSharedClient<fuchsia_wlan_wlanphyimpl::WlanphyImpl> client_;
    
      // Dispatcher for being a FIDL client firing requests to WlanphyImpl device.
      fdf::Dispatcher client_dispatcher_;
     ...
    

    (Source: device_dfv2.h)

    You can retrieve the driver's default dispatcher using the fdf::Dispatcher::GetCurrent() method or create a new, non-default dispatcher (see Update the DFv1 driver to use non-default dispatchers).

On the server end, do the following:

  1. Inherit a FIDL server class (fdf::WireServer<ProtocolName>) from the device class.

  2. Declare a dispatcher object (fdf::Dispatcher) that the FIDL server binds to.

    Unlike the client end, a dispatcher is always needed on the server end for binding the FIDL server object.

    The example code below shows the FIDL server and dispatcher objects in the device class:

    class Device : public fidl::WireServer<fuchsia_wlan_device::Phy>,
                   public ::ddk::Device<Device, ::ddk::MessageableManual, ::ddk::Unbindable> {
     ...
     private:
      // Dispatcher for being a FIDL server listening MLME requests.
      async_dispatcher_t* server_dispatcher_;
     ...
    

    (Source: device_dfv2.h)

    You can retrieve the driver's default dispatcher using the fdf::Dispatcher::GetCurrent() method or create a new, non-default dispatcher (see Update the DFv1 driver to use non-default dispatchers).

In the .fidl file, do the following:

  • Define a driver service protocol for the client and server ends.

    The example code below shows a driver service protocol defined in a .fidl file:

    service Service {
        wlan_phy_impl client_end:WlanPhyImpl;
    };
    

    (Source: phyimpl.fidl)

3. Update the driver to use the driver runtime FIDL

With the changes in the previous step, you can start updating your driver's implementation to use the driver runtime FIDL.

On the client end, do the following:

  1. To connect to the protocol that the parent device driver added to its outgoing directory, call the DdkConnectRuntimeProtocol() function, for example:

    auto client_end = DdkConnectRuntimeProtocol<fuchsia_wlan_softmac::Service::WlanSoftmac>();
    

    (Source: device.cc)

    This function creates a pair of endpoints:

    • The function returns the fdf::ClientEnd<ProtocolName> object to the caller.
    • The fdf::ServerEnd<ProtocolName> object silently goes to the parent device driver.
  2. When the caller gets the client end object, pass the object to the constructor of fdf::WireSharedClient<ProtocolName>() (or fdf::WireSyncClient<ProtocolName>()), for example:

    client_ = fdf::WireSharedClient<fuchsia_wlan_phyimpl::WlanPhyImpl>(std::move(client), client_dispatcher_.get());
    

    (Source: device.cc)

On the server end, do the following:

  1. Declare an outgoing directory object in the device class, for example:

    #include <lib/driver/outgoing/cpp/outgoing_directory.h>
    
    class Device : public DeviceType,
                   public fdf::WireServer<fuchsia_wlan_phyimpl::WlanPhyImpl>,
                   public DataPlaneIfc {
    ...
    
       fdf::OutgoingDirectory outgoing_dir_;
    

    (Source: device.h)

  2. Call the fdf::OutgoingDirectory::Create() function so that the parent driver creates an outgoing directory object, for example:

    #include <lib/driver/outgoing/cpp/outgoing_directory.h>
    
    ...
    
    Device::Device(zx_device_t *parent)
        : DeviceType(parent),
          outgoing_dir_(
              fdf::OutgoingDirectory::Create(
                 fdf::Dispatcher::GetCurrent()->get()))
    

    (Source: wlan_interface.cc)

  3. Add the service into the outgoing directory and serve it.

    The parent driver's service protocol is served to this outgoing directory so that the child driver can connect to it.

    In the parent driver's service callback function (which is a function that gets called when the child node is connected to the service), bind the fdf::ServerEnd<ProtocolName> object to itself using the fdf::BindServer() function, for example:

    zx_status_t Device::ServeWlanPhyImplProtocol(
            fidl::ServerEnd<fuchsia_io::Directory> server_end) {
      // This callback will be invoked when this service is being connected.
      auto protocol = [this](
          fdf::ServerEnd<fuchsia_wlan_phyimpl::WlanPhyImpl> server_end) mutable {
        fdf::BindServer(fidl_dispatcher_.get(), std::move(server_end), this);
        protocol_connected_.Signal();
      };
    
      // Register the callback to handler.
      fuchsia_wlan_phyimpl::Service::InstanceHandler handler(
           {.wlan_phy_impl = std::move(protocol)});
    
      // Add this service to the outgoing directory so that the child driver can
      // connect to by calling DdkConnectRuntimeProtocol().
      auto status =
           outgoing_dir_.AddService<fuchsia_wlan_phyimpl::Service>(
                std::move(handler));
      if (status.is_error()) {
        NXPF_ERR("%s(): Failed to add service to outgoing directory: %s\n",
             status.status_string());
        return status.error_value();
      }
    
      // Serve the outgoing directory to the entity that intends to open it, which
      // is DFv1 in this case.
      auto result = outgoing_dir_.Serve(std::move(server_end));
      if (result.is_error()) {
        NXPF_ERR("%s(): Failed to serve outgoing directory: %s\n",
             result.status_string());
        return result.error_value();
      }
    
      return ZX_OK;
    }
    

    (Source: device.cc)

    Notice that the fdf::BindServer() function requires a dispatcher as input. You can either use the default driver dispatcher provided by the driver host (fdf::Dispatcher::GetCurrent()->get()) or create a new, non-default dispatcher to handle FIDL requests separately (see Update the DFv1 driver to use non-default dispatchers).

    At this point, the client and server ends are ready to talk to each other using the driver runtime FIDL.

4. Make FIDL requests

For making FIDL calls, use the proxy of the fdf::WireSharedClient<ProtocolName>() object constructed on the client end from the previous steps.

See the syntax below for making FIDL calls (where CLIENT_ is the name of the instance):

  • Asynchronous FIDL call:

    CLIENT_.buffer(*std::move(arena))->MyExampleFunction().ThenExactlyOnce([](fdf::WireUnownedResult<FidlFunctionName>& result) mutable {
      // Your result handler.
    });
    

    (Source: device.cc)

  • Synchronous FIDL call:

    auto result = CLIENT_.sync().buffer(*std::move(arena))->MyExampleFunction();
    // Your result handler.
    

    (Source: device.cc)

You may find the following practices helpful for making FIDL calls:

  • Using FIDL error syntax, you can call result.is_error() to check whether the call returns a domain error. Similarly, the result.error_value() call returns the exact error value.

  • When making asynchronous two-way client FIDL calls, instead of passing a callback, you may use .Then(callback) or .ThenExactlyOnce(callback) to specify the desired cancellation semantics of pending callbacks.

  • Synchronous FIDL calls will wait for the callback from the server end. Therefore, a callback definition is required for this call in the .fidl file. Without it, the call will just do "fire and forget" and return immediately.

    For callback definitions, see the following example in a .fidl file:

    protocol MyExampleProtocol {
    // You can only make async calls based on this function.
      FunctionWithoutCallback();
    
      // You can make both sync and async calls based on this function.
      FunctionWithCallback() -> ();
    }
    

    Once the callback definitions are created, a function (which will be invoked in both synchronous and asynchronous callbacks in the example above) needs to be implemented on the server end (see MyExampleFunction() below).

    When the server end device class inherits the fdf::WireServer<ProtocolName> object, virtual functions based on your protocol definition are generated, similar to LLCPP. The following is the format of this function:

    void MyExampleFunction(MyExampleFunctionRequestView request, fdf::Arena& arena, MyExampleFunctionCompleter::Sync& completer);
    

    This function takes three parameters:

    • MyExampleFunctionRequestView – Generated by FIDL, it contains the request structure you want to send from the client to the server.

    • fdf::Arena – It is the buffer of this FIDL message. This buffer is passed or moved from the client end. You can use it as the buffer to return a message or error syntax through the completer object. You can also reuse it to make FIDL calls to next level drivers.

    • MyExampleFunctionCompleter – Generated by FIDL, it is used to invoke the callback and return the result of this FIDL call to the client end. If FIDL error syntax is defined, you can use completer.ReplySuccess() or completer.ReplyError() to return a message and error status, if FIDL error syntax is not defined, you can only use completer.Reply() to return the message.

  • You can either move the arena object into the FIDL call or simply pass it as *arena. However, moving the arena object may expose potential mistakes. Therefore, passing the arena object is recommended instead. Because the arena may be reused for the next level FIDL call, the arena object is not destroyed after being passed.

  • If your driver sends messages formatted as self-defined types in the .fidl file, FIDL generates both nature type and wire type based on the definition. However, the wire type is recommended in this case because the wire type is a more stable choice while the nature type is an interim type for HLCPP.

  • FIDL closes its channel if it notices that the endpoints exchanged invalid values, which is known as validation error:

    • A good example of a validation error is when passing 0 (for a non-flexible enum) while it only defines values 1 through 5. In case of a validation error, the channel is closed permanently and all subsequent messages result in failure. This behavior is different from Banjo. (For more information, see Strict vs. Flexible.)

    • Invalid values also include kernel objects with the zx_handle_t value set to 0. For example, zx::vmo(0).

(Optional) Update the DFv1 driver to use non-default dispatchers

The fdf::Dispatcher::GetCurrent() method gives you the default dispatcher that the driver is running on. If possible, it's recommended to use this default dispatcher alone. However, if you need to create a new, non-default dispatcher, the driver runtime needs to understand which driver the dispatcher belongs to. This allows the driver framework to set up proper attributes and shut down the dispatcher correctly.

The sections below describe advanced use cases of allocating and managing your own non-default dispatcher:

  1. Allocate a dispatcher.
  2. Shut down a dispatcher.

1. Allocate a dispatcher

To make sure that dispatchers are created and run in the threads managed by the driver framework, the allocation of dispatchers must be done in the functions invoked by the driver framework (such as DdkInit()), for example:

void Device::DdkInit(ddk::InitTxn txn) {
  bool fw_init_pending = false;
  const zx_status_t status = [&]() -> zx_status_t {
    auto dispatcher = fdf::SynchronizedDispatcher::Create(
        {}, "nxpfmac-sdio-wlanphy",
        [&](fdf_dispatcher_t *) { sync_completion_signal(&fidl_dispatcher_completion_); });
    if (dispatcher.is_error()) {
      NXPF_ERR("Failed to create fdf dispatcher: %s", dispatcher.status_string());
      return dispatcher.status_value();
    }
    fidl_dispatcher_ = std::move(*dispatcher);
  ...

(Source: device.cc)

2. Shut down a dispatcher

Likewise, the dispatcher shutdown also needs to happen in the functions invoked by the driver framework for the same reason. The DdkUnbind() or device_unbind() method is a good candidate for this operation.

Note that dispatcher shutdown is asynchronous, which needs to be handled appropriately. For example, if a dispatcher is shutting down in the DdkUnbind() call, we need to use the ddk::UnbindTxn object (passed from the Unbind() call previously) to invoke the ddk::UnbindTxn::Reply() call in the shutdown callback of the dispatcher to ensure a clean shutdown.

The following code snippet examples demonstrate the shutdown process described above:

  1. Save the ddk::UnbindTxn object in DdkUnbind():

    void DdkUnbind(ddk::UnbindTxn txn) {
      // Move the txn here because it’s not copyable.
      unbind_txn_ = std::move(txn);
      ...
    }
    
  2. As part of the bind or DdkInit() hook, create a dispatcher with a shutdown callback that invokes ddk::UnbindTxn::Reply():

      auto dispatcher = fdf::Dispatcher::Create(0, [&](fdf_dispatcher_t*) {
        if (unbind_txn_)
          unbind_txn_->Reply();
        unbind_txn_.reset();
      });
    
  3. Call ShutdownAsync() from the dispatcher at the end of DdkUnbind():

    void DdkUnbind(ddk::UnbindTxn txn) {
      // Move the txn here because it’s not copyable.
      unbind_txn_ = std::move(txn);
      ...
      dispatcher.ShutDownAsync();
    }
    

However, if there are more than one dispatcher allocated in the driver, because ddk::UnbindTxn::Reply() is called only once, you need to implement a chain of shutdown operations. For instance, given dispatchers A and B (which are interchangeable), you can:

  1. Call ShutdownAsync() for B in the shutdown callback of A.
  2. Call ddk::UnbindTxn::Reply() in the shutdown callback of B.

(Optional) Update the DFv1 driver to use two-way communication

It is common that only the device on the client end needs to proactively make requests to the server-end device. But in some cases, both devices need to send messages across the channel without having to wait for the other end to respond.

For establishing two-way communication in your DFv1 driver, you have the following three options:

  • Option 1 – Define events in the FIDL protocol (see Implement a C++ FIDL server).

  • Option 2 – Implement a second FIDL protocol in the opposite direction so that devices on both ends are both server and client at the same time.

  • Option 3 – Use the "hanging get" pattern in FIDL (a flow control design pattern recommended by the FIDL rubric).

Reasons to use events (option 1) include:

  • Simplicity - A single protocol is simpler than two.

  • Serialization - If you need two protocols, events and replies to requests are guaranteed to be serialized in the order they are written to the channel.

Reasons not to use events (option 1) include:

  • You need to respond to messages when they are sent from the server to the client.

  • You need to control the flow of messages.

The protocols below implement option 2 where they are defined in the same .fidl file for two different directions:

For the WLAN driver migration, the team selected option 2 since it wasn't going to introduce additional FIDL syntax from unknown domains. But it should be noted that the WLAN driver was the first instance of the driver runtime migration, and FIDL events in driver transport were not supported at the time of migration.

Depending on your needs, a "hanging get" call (option 3) is often a better option than an event because it allows the client to specify that it is ready to handle the event, which also provides flow control. (However, if necessary, there are alternative ways to add flow control to events, which are described in the Throttle events using acknowledgements section of the FIDL Rubric.)

Update the DFv1 driver's unit tests to use FIDL

If there exist unit tests based on the Banjo APIs for your driver, you need to migrate the tests to provide a mock FIDL server instead of the Banjo server (or a mock FIDL client).

In DFv1, to mock a FIDL server or client, you can use the MockDevice::FakeRootParent() method, for example:

std::shared_ptr<MockDevice> fake_parent_ = MockDevice::FakeRootParent();

(Source: ft_device_test.cc)

The MockDevice::FakeRootParent() method is integrated with the DriverRuntime testing library (which is the only supported testing library for DFv1). The MockDevice::FakeRootParent() method creates a fdf_testing::DriverRuntime instance for the user. The instance then starts the driver runtime and creates a foreground driver dispatcher for the user. However, background dispatchers can also be created through this object. You can grab the instance using the fdf_testing::DriverRuntime::GetInstance() method.

For an example, see this unit test that mocks the PWM and vreg FIDL protocols for a DFv1 driver.

Also, the following library may be helpful for writing driver unit tests:

Additional resources

All the Gerrit changes mentioned in this section:

All the source code files mentioned in this section:

All the documentation pages mentioned in this section: