Polling API

The polling API is used to wait concurrently for any one of multiple conditions to be fulfilled.

Concepts

The polling API’s main function is k_poll(), which is very similar in concept to the POSIX poll() function, except that it operates on kernel objects rather than on file descriptors.

The polling API allows a single thread to wait concurrently for one or more conditions to be fulfilled without actively looking at each one individually.

There is a limited set of such conditions:

  • a semaphore becomes available
  • a kernel FIFO contains data ready to be retrieved
  • a poll signal is raised

A thread that wants to wait on multiple conditions must define an array of poll events, one for each condition.

All events in the array must be initialized before the array can be polled on.

Each event must specify which type of condition must be satisfied so that its state is changed to signal the requested condition has been met.

Each event must specify what kernel object it wants the condition to be satisfied.

Each event must specify which mode of operation is used when the condition is satisfied.

Each event can optionally specify a tag to group multiple events together, to the user’s discretion.

Apart from the kernel objects, there is also a poll signal pseudo-object type that be directly signaled.

The k_poll() function returns as soon as one of the conditions it is waiting for is fulfilled. It is possible for more than one to be fulfilled when k_poll() returns, if they were fulfilled before k_poll() was called, or due to the preemptive multi-threading nature of the kernel. The caller must look at the state of all the poll events in the array to figured out which ones were fulfilled and what actions to take.

Currently, there is only one mode of operation available: the object is not acquired. As an example, this means that when k_poll() returns and the poll event states that the semaphore is available, the caller of k_poll() must then invoke k_sem_take() to take ownership of the semaphore. If the semaphore is contested, there is no guarantee that it will be still available when k_sem_give() is called.

Implementation

Using k_poll()

The main API is k_poll(), which operates on an array of poll events of type struct k_poll_event. Each entry in the array represents one event a call to k_poll() will wait for its condition to be fulfilled.

They can be initialized using either the runtime initializers K_POLL_EVENT_INITIALIZER() or k_poll_event_init(), or the static initializer K_POLL_EVENT_STATIC_INITIALIZER(). An object that matches the type specified must be passed to the initializers. The mode must be set to K_POLL_MODE_NOTIFY_ONLY. The state must be set to K_POLL_STATE_NOT_READY (the initializers take care of this). The user tag is optional and completely opaque to the API: it is there to help a user to group similar events together. Being optional, it is passed to the static initializer, but not the runtime ones for performance reasons. If using runtime initializers, the user must set it separately in the struct k_poll_event data structure. If an event in the array is to be ignored, most likely temporarily, its type can be set to K_POLL_TYPE_IGNORE.

struct k_poll_event events[2] = {
    K_POLL_EVENT_STATIC_INITIALIZER(K_POLL_TYPE_SEM_AVAILABLE,
                                    K_POLL_MODE_NOTIFY_ONLY,
                                    &my_sem, 0),
    K_POLL_EVENT_STATIC_INITIALIZER(K_POLL_TYPE_FIFO_DATA_AVAILABLE,
                                    K_POLL_MODE_NOTIFY_ONLY,
                                    &my_fifo, 0),
};

or at runtime

struct k_poll_event events[2];
void some_init(void)
{
    k_poll_event_init(&events[0],
                      K_POLL_TYPE_SEM_AVAILABLE,
                      K_POLL_MODE_NOTIFY_ONLY,
                      &my_sem);

    k_poll_event_init(&events[1],
                      K_POLL_TYPE_FIFO_DATA_AVAILABLE,
                      K_POLL_MODE_NOTIFY_ONLY,
                      &my_fifo);

    // tags are left uninitialized if unused
}

After the events are initialized, the array can be passed to k_poll(). A timeout can be specified to wait only for a specified amount of time, or the special values K_NO_WAIT and K_FOREVER to either not wait or wait until an event condition is satisfied and not sooner.

Only one thread can poll on a semaphore or a FIFO at a time. If a second thread tries to poll on the same semaphore or FIFO, k_poll() immediately returns with the return value -EADDRINUSE. In that case, if other conditions passed to k_poll() were met, their state will be set in the corresponding poll event.

In case of success, k_poll() returns 0. If it times out, it returns -EAGAIN.

// assume there is no contention on this semaphore and FIFO
// -EADDRINUSE will not occur; the semaphore and/or data will be available

void do_stuff(void)
{
    rc = k_poll(events, 2, 1000);
    if (rc == 0) {
        if (events[0].state == K_POLL_STATE_SEM_AVAILABLE) {
            k_sem_take(events[0].sem, 0);
        } else if (events[1].state == K_POLL_STATE_FIFO_DATA_AVAILABLE) {
            data = k_fifo_get(events[1].fifo, 0);
            // handle data
        }
    } else {
        // handle timeout
    }
}

When k_poll() is called in a loop, the events state must be reset to K_POLL_STATE_NOT_READY by the user.

void do_stuff(void)
{
    for(;;) {
        rc = k_poll(events, 2, K_FOREVER);
        if (events[0].state == K_POLL_STATE_SEM_AVAILABLE) {
            k_sem_take(events[0].sem, 0);
        } else if (events[1].state == K_POLL_STATE_FIFO_DATA_AVAILABLE) {
            data = k_fifo_get(events[1].fifo, 0);
            // handle data
        }
        events[0].state = K_POLL_STATE_NOT_READY;
        events[1].state = K_POLL_STATE_NOT_READY;
    }
}

Using k_poll_signal_raise()

One of the types of events is K_POLL_TYPE_SIGNAL: this is a “direct” signal to a poll event. This can be seen as a lightweight binary semaphore only one thread can wait for.

A poll signal is a separate object of type struct k_poll_signal that must be attached to a k_poll_event, similar to a semaphore or FIFO. It must first be initialized either via K_POLL_SIGNAL_INITIALIZER() or k_poll_signal_init().

struct k_poll_signal signal;
void do_stuff(void)
{
    k_poll_signal_init(&signal);
}

It is signaled via the k_poll_signal_raise() function. This function takes a user result parameter that is opaque to the API and can be used to pass extra information to the thread waiting on the event.

struct k_poll_signal signal;

// thread A
void do_stuff(void)
{
    k_poll_signal_init(&signal);

    struct k_poll_event events[1] = {
        K_POLL_EVENT_INITIALIZER(K_POLL_TYPE_SIGNAL,
                                 K_POLL_MODE_NOTIFY_ONLY,
                                 &signal),
    };

    k_poll(events, 1, K_FOREVER);

    if (events.signal->result == 0x1337) {
        // A-OK!
    } else {
        // weird error
    }
}

// thread B
void signal_do_stuff(void)
{
    k_poll_signal_raise(&signal, 0x1337);
}

If the signal is to be polled in a loop, both its event state and its signaled field must be reset on each iteration if it has been signaled.

struct k_poll_signal signal;
void do_stuff(void)
{
    k_poll_signal_init(&signal);

    struct k_poll_event events[1] = {
        K_POLL_EVENT_INITIALIZER(K_POLL_TYPE_SIGNAL,
                                 K_POLL_MODE_NOTIFY_ONLY,
                                 &signal),
    };

    for (;;) {
        k_poll(events, 1, K_FOREVER);

        if (events[0].signal->result == 0x1337) {
            // A-OK!
        } else {
            // weird error
        }

        events[0].signal->signaled = 0;
        events[0].state = K_POLL_STATE_NOT_READY;
    }
}

Suggested Uses

Use k_poll() to consolidate multiple threads that would be pending on one object each, saving possibly large amounts of stack space.

Use a poll signal as a lightweight binary semaphore if only one thread pends on it.

Note

Because objects are only signaled if no other thread is waiting for them to become available and only one thread can poll on a specific object, polling is best used when objects are not subject of contention between multiple threads, basically when a single thread operates as a main “server” or “dispatcher” for multiple objects and is the only one trying to acquire these objects.

Configuration Options

Related configuration options: