Fibers¶

lbuild module: modm:processing:fiber

This module provides a lightweight stackful fiber implementation including a simple round-robin scheduler. Here is a minimal example that blinks an LED:

modm::Fiber<> fiber([]
{
    Board::LedBlue::setOutput();
    while(true)
    {
        Board::LedBlue::toggle();
        modm::this_fiber::sleep_for(1s);
    }
});
int main()
{
    modm::fiber::Scheduler::run();
    return 0;
}

Construction¶

You can construct a fiber from any function without return type or arguments:

modm::Fiber<> fiber([]{});
void function() {}
modm::Fiber<> fiber2(function);

To call objects with arguments, wrap the data into a lambda closure and pass it to fiber. The closure will be constructed at the top of the stack and allows the lambda wrapper to call your function with an argument:

struct DataObject
{
    void member_function(int arg);
} object;
int number{42};
modm::Fiber<> fiber([&]
{
    object.member_function(number);
});

Remember to use the right capture method for the lifetime of the objects you want to call. You can std::move() already constructed objects into the capture, or construct them in the capture directly, if they would get destroyed after fiber construction. You may need to mark the lambda mutable:

modm::Fiber<> fiber2([obj=std::move(object), obj2=DataObject()] mutable
{
    obj.member_function(24);
    obj2.member_function(42);
});

Do not construct modm::Fiber on the stack!

Apart from the general lifetime issues of constructing objects on the stack, the allocated fiber stack size is likely too large for the caller stack and will lead to a stack overflow.

A fiber can be passed a modm::fiber::stop_token to allow the fiber to be stopped cooperatively.

modm::Fiber<> fiber([](modm::fiber::stop_token stoken)
{
    // set up
    while(not stoken.stop_requested())
    {
        // run your task
    }
    // clean up
});
// externally request the fiber to stop
fiber.request_stop();
// wait until fiber has stopped
fiber.join();

Note that the fiber destructor requests to stop and joins automatically. The interface and behavior is similar to the C++20 std::jthread.

Delayed Start¶

Fiber are added to the scheduler automatically and start execution when the scheduler is run. You can disable this behavior by setting start to modm::fiber::Start::Later during construction and manually starting the fiber when it is ready, also from another fiber:

// fiber does not automatically start executing
modm::Fiber<> fiber2(function, modm::fiber::Start::Later);
// fiber2 is automatically executing
modm::Fiber<> fiber1([&]
{
    modm::this_fiber::sleep_for(1s);
    fiber2.start();
});
modm::fiber::Scheduler::run();
// fiber1 waits 1s, then starts fiber2 and exits

Fibers can end by returning from their wrapper, after which they will be removed from the scheduler. A fiber can then be restarted again by calling start(), which will call the closure again from the beginning. Note, that the lambda capture is not destructed and reconstructed, but remains unchanged between restarts. If you need a fiber that is only callable once, you can implement this behavior manually with a boolean in the capture:

modm::Fiber<> fiber([ran=false]
{
    if (ran) return;
    ran = true;
    // only called once.
});

Customization¶

The most important customization is the fiber stack size expressed in bytes:

modm::Fiber<128> fiber1(...);
modm::Fiber<256> fiber2(...);

The Fiber class is intentionally constructed at runtime, so that it does not increase your program size, as the .data section would. You may also place the fibers into the .faststack section, which is not zeroed and thus saves a bit of time on startup:

modm_faststack modm::Fiber<>(stack, function);

However, it may be desirable to control the placement of the fiber task structure and especially the stack, depending on the types of memories available on your device. This is possible when you construct the stack and task in combination with the modm_section() macros and its specializations:

// Place a very large stack in the external memory
modm_section(".sdram_noinit") modm::fiber::Stack<1024*10> large_stack;
// But keep the task control structure in internal memory
modm_fastdata modm::fiber::Task fiber(large_stack, big_function);

Subclassing¶

You can inherit from the modm::Fiber class to extend a fiber with a proper object interface. This is useful for providing a more expressive interface to a fiber by means of shared memory.

class CustomFiber : public modm::Fiber<>
{
public:
    CustomFiber() : Fiber([this]{ run(); }) {}

    // called from another fiber
    void trigger() { trigger_ = true; }

private:
    bool trigger_{false}; // shared memory between fibers

    void inline
    run()
    {
        while(true)
        {
            modm::this_fiber::poll([&]{ return trigger_; });
            trigger_ = false;
            // long running response to the trigger
        }
    }
};

Concurrency Support¶

The modm::fiber namespace provides several standard concurrency primitives to synchronize fibers based on the std::thread interface behavior. Most primitives are implemented on top of <atomic>, therefore can be called from within (nested) interrupts. The API docs explicitly mention if a function is safe to call from an interrupt.

Threads¶

Task implements most of the std::jthread interface.

In particular, Task only implements functionality that does not require dynamic memory allocations. The stack memory needs to be allocated externally and fibers are not movable or copyable and therefore cannot be detached or swapped.

Thread Cancellation¶

stop_token and stop_source with simplified implementations.
stop_callback not implemented.

To avoid dynamic memory allocations, a stop_state object provides the actual memory required for the limited functionality:

modm::fiber::stop_state state;
// only valid as long as state is valid!
auto source = state.get_source();
auto token = state.get_token();
// use token in a condition variable
cv.wait(lock, token, predicate);
// request a stop somewhere else
source.request_stop();

Implemented using interrupt-safe atomics.

Mutual Exclusion¶

mutex and timed_mutex.
recursive_mutex and recursive_timed_mutex.
shared_mutex and shared_timed_mutex.

Implemented using interrupt-safe atomics.

Generic Mutex Management¶

lock_guard, scoped_lock, unique_lock and shared_lock.
defer_lock_t, try_to_lock_t and adopt_lock_t.
defer_lock, try_to_lock and adopt_lock.

Generic Locking Algorithms¶

try_lock and lock.

Call Once¶

once_flag and call_once.

Implemented using interrupt-safe atomic flag.

Condition Variables¶

condition_variable and condition_variable_any.
cv_status.
notify_all_at_thread_exit not implemented.

Notification is implemented as a interrupt-safe 16-bit atomic counter.

Semaphores¶

counting_semaphore and binary_semaphore.

Counts are implemented as interrupt-safe 16-bits atomics.

Latches and Barriers¶

latch: implemented as interrupt-safe atomics.
barrier: not interrupt-safe!

Counts are implemented as 16-bits.

Stack Usage¶

To measure the stack usage of a fiber, you need to explicitly watermark the stack before running the fiber, then you may query the stack usage inside or outside the fiber:

// You must watermark the stack *before* running the fiber!
fiber1.stack_watermark();
// now you can run the fibers via the scheduler
modm::fiber::Scheduler::run();
// can be called from inside or outside the fiber, before or after running!
size_t bytes = fiber.stack_usage();

Note that stack usage measurement through watermarking can be inaccurate if the registers contain the watermark value.

Stack Overflow¶

Each context switch checks if the stack overflowed, in which case the scheduler will abandon execution and trigger an assertion on the main stack with the identifier fbr.stkof and the fiber pointer as context. Note that the assertion is executed on the main stack and not on the fiber stack that overflowed!

On ARMv8-M devices, the stack overflow is checked in hardware via the PSPLIM register, therefore the context switch is a little faster.

Scheduling¶

The scheduler run() function will suspend execution of the call site, usually the main function, start each fiber and continue to execute them until they all ended and then return execution to the call site:

while(true)
{
    modm::fiber::Scheduler::run();
    // sleep until the next interrupt?
    __WFI();
    // then start the fibers again
    fiber.start();
}

Please note that neither the fiber nor scheduler is interrupt safe, so starting threads from interrupt context is a bad idea!

Using yield() outside of a fiber

If yield() is called before the scheduler started or if only one fiber is running, it simply returns in-place, since there is nowhere to switch to.

Platforms¶

Fibers are implemented by saving callee registers to the current stack, then switching to a new stack and restoring callee registers from this stack. The static modm::this_fiber::yield() function wraps this functionality in a transparent way.

AVR¶

On AVRs the fiber stack is shared with the currently active interrupt. This requires the fiber stack size to include the worst case stack size of all interrupts. Fortunately on AVRs interrupts cannot be nested.

Therefore the default stack size is a fairly large 512B.

Arm Cortex-M¶

On Cortex-M, the main function is entered using the MSP in Handler mode. After calling modm::fiber::Scheduler::run() the PSP is used as a Fiber stack pointer in Thread mode. Therefore all interrupts are using the main stack whose size is defined by the modm:platform:cortex-m:main_stack_size option and will not increase the fiber stack size at all.

The default stack size is 1KiB.

Hosted¶

Two implementations for x86_64 and ARM64 are provided.

The default stack size is 1MiB.

Multi-Core Scheduling¶

When using this module in combination with the modm:platform:multicore module, each core gets its own fiber scheduler, which will internally be selected based on the CPU ID. Since the scheduler is not thread-safe, you cannot add fibers from one core to the other. Instead you must construct the fiber without starting it, and when executing on the other core, start() it in that context.

Here is an example for the RP2040 device, which additionally allocates the stack and task into the core-affine memory:

// allocate into core0 memory
modm_faststack_core0 modm::Fiber<> fiber0(function);
// allocate into core1 memory but DO NOT start yet!
modm_faststack_core1 modm::Fiber<> fiber1(function, modm::fiber::Start::Later);

void core1_main()
{
    // start fiber1 in the core1 context!
    fiber1.start();
    modm::fiber::Scheduler::run();
}

int main()
{
    modm::platform::multicore::Core1::run(core1_main);
    // run fiber0 in core0 context
    modm::fiber::Scheduler::run();
    return 0;
}