BloomPatched/src/TargetController

TargetController

The TargetController component possesses full control of the connected hardware (debug tool and target). Execution of user-space device drivers takes place here. All interactions with the connected hardware go through the TargetController. It runs on a dedicated thread (see Applciation::startTargetController()). The source code for the TargetController component can be found in src/TargetController. The entry point is TargetControllerComponent::run().

Lifetime

The TargetController is the second (after the SignalHandler) component to be instantiated at startup, and the second to last to be destroyed when Bloom shuts down.

The TargetController will outlive both the DebugServer and the Insight GUI components.

Interfacing with the TargetController - The command-response mechanism

Other components within Bloom can interface with the TargetController via the provided command-response mechanism. Simply put, when another component within Bloom needs to interact with the connected hardware, it will send a command to the TargetController, and wait for a response. The TargetController will action the command and deliver the necessary response.

All TargetController commands can be found in src/TargetController/Commands, and are derived from the TargetController::Commands::Command base class. Responses can be found in src/TargetController/Responses, and are derived from the TargetController::Responses::Response base class.

NOTE: Components within Bloom do not typically concern themselves with the TargetController command-response mechanism. Instead, they use the TargetControllerService class, which encapsulates the command-response mechanism and provides a simplified means for interaction with the connected hardware. For more, see The TargetControllerService class section below.

Commands can be sent to the TargetController via the TargetController::CommandManager class.

For example, to read memory from the connected target, we would send the TargetController::Commands::ReadTargetMemory command:

auto tcCommandManager = TargetController::CommandManager{};

auto readMemoryCommand = std::make_unique<TargetController::Commands::ReadTargetMemory>(
    addressSpaceDescriptor,
    memorySegmentDecriptor,
    someStartAddress,
    bytesToRead
);

const auto readMemoryResponse = tcCommandManager.sendCommandAndWaitForResponse(
    std::move(readMemoryCommand),
    std::chrono::milliseconds{1000} // Response timeout
);

const auto& data = readMemoryResponse->data;

readMemoryResponse will be of type std::unique_ptr<TargetController::Responses::TargetMemoryRead>.

The CommandManager::sendCommandAndWaitForResponse<CommandType>(std::unique_ptr<CommandType> command, std::chrono::milliseconds timeout) member function is a template function. It will issue the command to the TargetController and wait for a response, or until a timeout has been reached. Because it is a template function, it is able to resolve the appropriate response type at compile-time (see the SuccessResponseType alias in some command classes). If the TargetController responds with an error, or the timeout is reached, CommandManager::sendCommandAndWaitForResponse() will throw an exception.

The TargetControllerService class

The TargetControllerService class encapsulates the TargetController's command-response mechanism and provides a simplified means for other components to interact with the connected hardware. Iterating on the example above, to read memory from the target:

const auto tcService = Services::TargetControllerService{};

const auto data = tcService.readMemory(
    addressSpaceDescriptor,
    memorySegmentDecriptor,
    someStartAddress,
    bytesToRead
);

The TargetControllerService class does not require any dependencies at construction. It can be constructed in different threads and used freely to gain access to the connected hardware, from any component within Bloom.

All components within Bloom should use the TargetControllerService class to interact with the connected hardware. They should not directly issue commands via the TargetController::CommandManager, unless there is a very good reason to do so.

Atomic operations

In some instances, we need the TargetController to service a series of commands without any interruptions (servicing of other commands).

The TargetController allows for operations to be performed within "atomic sessions". Simply put, when the TargetController starts a new atomic session, any commands that are part of the session will be placed into a dedicated queue. When an atomic session is active, the TargetController will only process commands in the dedicated queue. All other commands will be processed once the atomic session has ended.

Atomic sessions with the TargetControllerService

The TargetControllerService provides an RAII wrapper for atomic sessions.

The TargetControllerService::makeAtomicSession() member function returns an TargetControllerService::AtomicSession RAII object, which starts an atomic session with the TargetController, at construction, and ends the session at destruction. This allows us to perform operations within an atomic session, in an exception-safe manner:

auto tcService = Services::TargetControllerService{};

{
    const auto atomicSession = tcService.makeAtomicSession();

    /*
     * These operations will take place in the atomic session - the TC will **NOT** service any other commands until
     * these commands have been processed (and the atomic session has ended).
     */
    tcService.writeMemory(...);
    tcService.readMemory(...);
    tcService.getProgramCounter();

    /*
     * Note: The TC does **NOT** support nested atomic sessions. Attempting to start another session in this block will
     * result in an exception being thrown.
     */
    {
        const auto nestedAtomicSession = tcService.makeAtomicSession(); // This will fail - a session is already active.
    }

    /*
     * Also note: When using TargetControllerService::makeAtomicSession(), the returned AtomicSession object is tied to
     * the TargetControllerService object that created it (in this example: tcService).
     *
     * So if you have **another** TargetControllerService object in this block, any operations performed via that
     * object will **NOT** be part of the atomic session, and, they will deadlock the TC. So don't ever do this.
     * You should never need more than one TargetControllerService object in a single block.
     */

    // Don't ever do this.
    auto anotherTcService = Services::TargetControllerService{};

    // These operations will **NOT** be part of the atomic session, and they will cause a deadlock and timeout.
    anotherTcService.writeMemory(...);
    anotherTcService.readMemory(...);

    /*
     * One more thing: The AtomicSession object should **NEVER** outlive the TargetControllerService object that
     * created it.
     *
     * If this happens, the AtomicSession will have a dangling reference, which will result in UB.
     */
}

// At this point, the atomic session will have ended. The TC will now process any other commands in the queue.
tcService.readMemory(...); // Will not be part of the atomic session

Obtaining information on the connected target - target descriptor objects

Most components in Bloom require access to some target information. This information could be related to address spaces, peripherals, registers, or any other target entity.

In Bloom, target information is held in descriptor objects. These objects are instances of descriptor structs intended to hold information on a particular target entity. For example, information on a particular address space can be found in the relevant TargetAddressSpaceDescriptor object.

All descriptor structs can be found in src/Targets/.

The TargetDescriptor struct is the top-level descriptor struct. Instances of all other descriptor structs can be accessed via the TargetDescriptor struct.

The GetTargetDescriptor command returns a const reference to the top-level target descriptor object. The TargetControllerService::getTargetDescriptor() member function can be used to obtain the top-level target descriptor:

const auto& targetDescriptor = tcService.getTargetDescriptor();

// Obtain info on the target's "data" address space
const auto& dataAsDescriptor = targetDescriptor.getAddressSpaceDescriptor("data");
/**
 * - dataAsDescriptor.addressRange.startAddress <- the start address of the address space
 * - dataAsDescriptor.size() <- the size of the address space, in bytes
 * - dataAsDescriptor.endianness <- the endianness of the address space
 *
 * See the TargetAddressSpaceDescriptor struct for more...
 */

// Obtain info on the target's "ADC0" peripheral
const auto& adcPerDescriptor = targetDescriptor.getPeripheralDescriptor("adc0");
/**
 * - adcPerDescriptor.name <- the name of the peripheral
 * - adcPerDescriptor.getRegisterGroupDescriptor(...) <- Access a register group within the peripheral
 *
 * See the TargetPeripheralDescriptor struct for more...
 */

The copying of descriptor objects is discouraged for two reasons:

• Descriptor objects hold a lot of information, making them expensive to copy
• Descriptor objects can be passed back to the TargetController, in the form of const references, when performing an operation on a particular target entity. If the copied object does not outlive the TargetController command, the TargetController may attempt to access a dangling reference, resulting in undefined behaviour. For more on this, see the Passing descriptor objects to the TargetController section below

The copy constructors and the assignment operators are deleted on all descriptor structs.

Passing descriptor objects to the TargetController

For some operations, the TargetController may require the referencing of a particular target entity. For example, to read memory from the target, the TargetController will need to know which address space and memory segment to read from. This is why the ReadTargetMemory command requires a const TargetAddressSpaceDescriptor& and const TargetMemorySegmentDescriptor&. There are many commands that require const references to target descriptor objects.

It is the invoker's responsibility to ensure that all const references to descriptor objects, passed to the TargetController, are still valid at the time the TargetController services the command. If the TargetController encounters a dangling reference, undefined behaviour will occur. See the "master target descriptor" for more.

Thread-safe access of the target descriptor - the master target descriptor object

The "master" target descriptor object is simply an instance of the top-level TargetDescriptor struct, owned and managed by the TargetController. The TargetController will return a const reference of the master target descriptor when servicing the GetTargetDescriptor command.

The TargetController will never modify the master target descriptor after creation. There are no mutable members in any of the target descriptor structs. This means we can access a const reference of the descriptor, from any thread within Bloom, assuming the thread does not outlive the TargetController.

Because the master target descriptor is owned by the TargetController, it, along with any of its contained objects, can be passed safely to the TargetController, in any TargetController command.

Both the debug server and Insight GUI components hold a const reference of the master target descriptor, allowing them to access target information freely.

Do not modify the master target descriptor. Do not perform a const_cast on the const reference. Ever.

Target state observation

The TargetController provides access to the target's current state, via the GetTargetState command, which will return an instance of the TargetState struct. This struct holds the execution state (TargetExecutionState), the mode (programming/debugging, TargetMode), and the program counter.

The TargetControllerService::getTargetState() member function should be used to obtain the target's current state:

const auto targetState = tcService.getTargetState();

if (targetState.executionState == TargetExecutionState::STOPPED) {
    // ...
}

Real-time, on-demand, thread-safe access to the target's current state - the master target state object

All members of the TargetState struct are accessible via atomic operations (that is, all members are of std::atomic<...> type). This means that we can access a single instance of the TargetState struct across multiple threads, in a thread-safe manner.

The "master" target state object is simply an instance of the TargetState struct that is owned and managed by the TargetController (TargetControllerComponent::targetState). It holds the current state of the target, at all times.

When servicing the GetTargetState command, the TargetController returns a const reference to the master target state object. This means that, if the caller of TargetControllerService::getTargetState() needs real-time, on-demand access to the target's current state, it can gain this by simply accepting a const reference of the master target state object:

const auto& targetState = tcService.getTargetState();

/*
 * In the previous example, we used `const auto targetState = tcService.getTargetState();`, which made a copy of the
 * master TargetState object. That copy would not be managed by the TargetController, and would only hold the state of
 * the target at the point when `tcService.getTargetState()` returned a value.
 *
 * In this example, `targetState` is a const reference to the master TargetState object - it will always hold the
 * target's current state.
 *
 * We can now observe the target's current state, without having to make any more calls to `TargetControllerService::getTargetState()`.
 */

if (targetState.executionState == TargetExecutionState::STOPPED) {
    tcService.resumeTargetExecution();

    /*
     * At this point, targetState.executionState == TargetExecutionState::RUNNING, because the master target state
     * object, which `targetState` references, will have been updated by the TargetController (as a result of the call
     * to `tcService.resumeTargetExecution()` above).
     *
     * This, of course, assumes that the target is still running by the time we get to this assertion...
     */
    assert(targetState.executionState == TargetExecutionState::RUNNING);
}

The master target state object can be accessed freely by any other component within Bloom, just as long as the component doesn't outlive the TargetController.

Many Insight GUI widgets make use of the master target state object, as it allows for immediate access to the target's current state, without having to bother the TargetController via an InsightWorker task.

Target state changed events

When the target state changes, the TargetController will emit a TargetStateChanged event. The event object holds two TargetState objects: TargetStateChanged::newState and TargetStateChanged::previousState. Listeners can use these to determine what changed:

void onTargetStateChanged(const Events::TargetStateChanged& event) {
    using Targets::TargetExecutionState;

    if (event.previousState.executionState == event.newState.executionState) {
        // Target execution state has not changed. Probably a mode change
    }

    if (
        event.previousState.executionState == TargetExecutionState::STOPPED
        && event.newState.executionState == TargetExecutionState::RUNNING
    ) {
        // Target has just resumed execution...
    }

    if (
        event.previousState.executionState == TargetExecutionState::STEPPING
        && event.newState.executionState == TargetExecutionState::STOPPED
    ) {
        // Target has just finished stepping...
    }

    // ...
}

Programming mode

When a component needs to write to the target's program memory, it must enable programming mode on the target. This can be done by issuing the EnableProgrammingMode command to the TargetController (see TargetControllerService::enableProgrammingMode()).

Once programming mode has been enabled, standard debugging operations such as program flow control and RAM access will become unavailable. The TargetController will reject any commands that involve these operations, until programming mode has been disabled. The Command::requiresDebugMode() virtual member function communicates a particular command's requirement for the target to not be in programming mode.

For example, the ResumeTargetExecution command returns true here, as it attempts to control program flow on the target, which can only be done when the target is not in programming mode:

class ResumeTargetExecution: public Command
{
public:
    // ...
    [[nodiscard]] bool requiresDebugMode() const override {
        return true;
    }
};

On the other hand, the ReadTargetMemory command will only return true if we're reading from RAM, as RAM is the only memory which isn't accessible when the target is in programming mode:

class ReadTargetMemory: public Command
{
public:
    Targets::TargetMemoryType memoryType;
    Targets::TargetMemoryAddress startAddress;
    Targets::TargetMemorySize bytes;

    // ...

    [[nodiscard]] bool requiresDebugMode() const override {
        return this->memoryType == Targets::TargetMemoryType::RAM;
    }
};

The TargetController will emit ProgrammingModeEnabled and ProgrammingModeDisabled events when it enables/disables programming mode. Components should listen for these events to ensure that they disable any means for the user to trigger debugging operations whilst programming mode is enabled. For example, the Insight component will disable much of its GUI components when programming mode is enabled.

It shouldn't be too much of a problem if a component attempts to perform a debug operation on the target whilst programming mode is enabled, as the TargetController will just respond with an error. But still, it would be best to avoid doing this where possible.

---

TODO: Cover debug tool & target drivers.