lib/isc/task.c

	task.c revision 0c7c25a1c64a2224053a1e3542211f2105e40fc8

#include "attribute.h"

#include <isc/assertions.h>

#include <isc/thread.h>
#include <isc/mutex.h>
#include <isc/condition.h>

#include <isc/task.h>


/***
 *** General Macros.
 ***/

/*
 * We use macros instead of calling the os_ routines directly because
 * the capital letters make the locking stand out.
 *
 * We INSIST that they succeed since there's no way for us to continue
 * if they fail.
 */
#define LOCK(lp)        INSIST(os_mutex_lock((lp)))
#define UNLOCK(lp)      INSIST(os_mutex_unlock((lp)))
#define WAIT(cvp, lp)       INSIST(os_condition_wait((cvp), (lp)))
#define BROADCAST(cvp)      INSIST(os_condition_broadcast((cvp)))

#ifdef TASK_TRACE
#define XTRACE(m)       printf("%s task %p thread %p\n", (m), \
                       task, os_thread_self())
#else
#define XTRACE(m)
#endif


/***
 *** Types.
 ***/

typedef enum {
    task_state_idle, task_state_ready, task_state_running,
    task_state_shutdown
} task_state_t;

#define TASK_MAGIC          0x5441534BU /* TASK. */
#define VALID_TASK(t)           ((t) != NULL && \
                     (t)->magic == TASK_MAGIC)

struct task {
    /* Not locked. */
    unsigned int            magic;
    task_manager_t          manager;
    os_mutex_t          lock;
    /* Locked by task lock. */
    task_state_t            state;
    unsigned int            references;
    task_eventlist_t        events;
    unsigned int            quantum;
    boolean_t           enqueue_allowed;
    task_event_t            shutdown_event;
    /* Locked by task manager lock. */
    LINK(struct task)       link;
    LINK(struct task)       ready_link;
};

#define TASK_MANAGER_MAGIC      0x54534B4DU /* TSKM. */
#define VALID_MANAGER(m)        ((m) != NULL && \
                     (m)->magic == TASK_MANAGER_MAGIC)

struct task_manager {
    /* Not locked. */
    unsigned int            magic;
    mem_context_t           mctx;
    os_mutex_t          lock;
    /* Locked by task manager lock. */
    unsigned int            default_quantum;
    LIST(struct task)       tasks;
    LIST(struct task)       ready_tasks;
    os_condition_t          work_available;
    boolean_t           exiting;
    unsigned int            workers;
    os_condition_t          no_workers;
};

#define DEFAULT_DEFAULT_QUANTUM     5
#define FINISHED(m)         ((m)->exiting && EMPTY((m)->tasks))


/***
 *** Events.
 ***/

static inline task_event_t
event_allocate(mem_context_t mctx, void *sender, task_eventtype_t type,
           task_action_t action, void *arg, size_t size)
{
    task_event_t event;

    event = mem_get(mctx, size);
    if (event == NULL)
        return (NULL);
    event->mctx = mctx;
    event->size = size;
    event->sender = sender;
    event->type = type;
    event->action = action;
    event->arg = arg;

    return (event);
}

task_event_t
task_event_allocate(mem_context_t mctx, void *sender, task_eventtype_t type,
            task_action_t action, void *arg, size_t size)
{
    if (size < sizeof (struct task_event))
        return (NULL);
    if (type < 0)
        return (NULL);
    if (action == NULL)
        return (NULL);

    return (event_allocate(mctx, sender, type, action, arg, size));
}

void
task_event_free(task_event_t *eventp) {
    task_event_t event;

    REQUIRE(eventp != NULL);
    event = *eventp;
    REQUIRE(event != NULL);

    mem_put(event->mctx, event, event->size);

    *eventp = NULL;
}

/***
 *** Tasks.
 ***/

static void
task_free(task_t task) {
    task_manager_t manager = task->manager;

    XTRACE("free task");
    REQUIRE(EMPTY(task->events));

    LOCK(&manager->lock);
    UNLINK(manager->tasks, task, link);
    if (FINISHED(manager)) {
        /*
         * All tasks have completed and the
         * task manager is exiting.  Wake up
         * any idle worker threads so they
         * can exit.
         */
        BROADCAST(&manager->work_available);
    }
    UNLOCK(&manager->lock);
    (void)os_mutex_destroy(&task->lock);
    if (task->shutdown_event != NULL)
        task_event_free(&task->shutdown_event);
    task->magic = 0;
    mem_put(manager->mctx, task, sizeof *task);
}

boolean_t
task_create(task_manager_t manager, task_action_t shutdown_action,
        void *shutdown_arg, unsigned int quantum, task_t *taskp)
{
    task_t task;

    REQUIRE(VALID_MANAGER(manager));
    REQUIRE(taskp != NULL && *taskp == NULL);

    task = mem_get(manager->mctx, sizeof *task);
    if (task == NULL)
        return (FALSE);

    task->magic = TASK_MAGIC;
    task->manager = manager;
    if (!os_mutex_init(&task->lock)) {
        mem_put(manager->mctx, task, sizeof *task);
        return (FALSE);
    }
    task->state = task_state_idle;
    task->references = 1;
    INIT_LIST(task->events);
    task->quantum = quantum;
    task->enqueue_allowed = TRUE;
    task->shutdown_event = event_allocate(manager->mctx,
                          NULL,
                          TASK_EVENT_SHUTDOWN,
                          shutdown_action,
                          shutdown_arg,
                          sizeof *task->shutdown_event);
    if (task->shutdown_event == NULL) {
        (void)os_mutex_destroy(&task->lock);
        mem_put(manager->mctx, task, sizeof *task);
        return (FALSE);
    }
    INIT_LINK(task, link);
    INIT_LINK(task, ready_link);

    LOCK(&manager->lock);
    if (task->quantum == 0)
        task->quantum = manager->default_quantum;
    APPEND(manager->tasks, task, link);
    UNLOCK(&manager->lock);

    *taskp = task;

    return (TRUE);
}

void
task_attach(task_t task, task_t *taskp) {

    REQUIRE(VALID_TASK(task));
    REQUIRE(taskp != NULL && *taskp == NULL);

    LOCK(&task->lock);
    task->references++;
    UNLOCK(&task->lock);

    *taskp = task;
}

void
task_detach(task_t *taskp) {
    boolean_t free_task = FALSE;
    task_t task;

    XTRACE("task_detach");

    REQUIRE(taskp != NULL);
    task = *taskp;
    REQUIRE(VALID_TASK(task));

    LOCK(&task->lock);
    REQUIRE(task->references > 0);
    task->references--;
    if (task->state == task_state_shutdown && task->references == 0)
        free_task = TRUE;
    UNLOCK(&task->lock);

    if (free_task)
        task_free(task);

    *taskp = NULL;
}

boolean_t
task_send_event(task_t task, task_event_t *eventp) {
    boolean_t was_idle = FALSE;
    boolean_t discard = FALSE;
    task_event_t event;

    REQUIRE(VALID_TASK(task));
    REQUIRE(eventp != NULL);
    event = *eventp;
    REQUIRE(event != NULL);
    REQUIRE(event->sender != NULL);
    REQUIRE(event->type > 0);

    XTRACE("sending");
    /*
     * We're trying hard to hold locks for as short a time as possible.
     * We're also trying to hold as few locks as possible.  This is why
     * some processing is deferred until after a lock is released.
     */
    LOCK(&task->lock);
    if (task->enqueue_allowed) {
        if (task->state == task_state_idle) {
            was_idle = TRUE;
            INSIST(EMPTY(task->events));
            task->state = task_state_ready;
        }
        INSIST(task->state == task_state_ready ||
               task->state == task_state_running);
        ENQUEUE(task->events, event, link);
    } else
        discard = TRUE;
    UNLOCK(&task->lock);

    if (discard) {
        task_event_free(&event);
        *eventp = NULL;
        return (TRUE);
    }

    if (was_idle) {
        boolean_t need_wakeup = FALSE;
        task_manager_t manager;

        /*
         * We need to add this task to the ready queue.
         *
         * We've waited until now to do it, rather than doing it
         * while holding the task lock, because we don't want to
         * block while holding the task lock.
         *
         * We've changed the state to ready, so no one else will
         * be trying to add this task to the ready queue.  It
         * thus doesn't matter if more events have been added to
         * the queue after we gave up the task lock.
         *
         * Shutting down a task requires posting a shutdown event
         * to the task's queue and then executing it, so there's
         * no way the task can disappear.  A task is always on the
         * task manager's 'tasks' list, so the task manager can
         * always post a shutdown event to all tasks if it is
         * requested to shutdown.
         */
        manager = task->manager;
        INSIST(VALID_MANAGER(manager));
        LOCK(&manager->lock);
        if (EMPTY(manager->ready_tasks))
            need_wakeup = TRUE;
        ENQUEUE(manager->ready_tasks, task, ready_link);
        UNLOCK(&manager->lock);

        /*
         * If the runnable queue is empty, the worker threads could
         * either be executing tasks or waiting for something to do.
         * We wakeup anyone who is sleeping.
         */
        if (need_wakeup)
            BROADCAST(&manager->work_available);
    }

    *eventp = NULL;

    XTRACE("sent");
    return (TRUE);
}

void
task_purge_events(task_t task, void *sender, task_eventtype_t type) {
    task_event_t event, next_event;
    task_eventlist_t purgeable;

    REQUIRE(VALID_TASK(task));
    REQUIRE(type >= 0);

    /*
     * Purge events matching 'sender' and 'type'.  sender == NULL means
     * "any sender".  type == NULL means any type.  Task manager events
     * cannot be purged.
     */

    INIT_LIST(purgeable);

    LOCK(&task->lock);
    for (event = HEAD(task->events);
         event != NULL;
         event = next_event) {
        next_event = NEXT(event, link);
        if ((sender == NULL || event->sender == sender) &&
            ((type == 0 && event->type > 0) || event->type == type)) {
            DEQUEUE(task->events, event, link);
            ENQUEUE(purgeable, event, link);
        }
    }
    UNLOCK(&task->lock);

    for (event = HEAD(purgeable);
         event != NULL;
         event = next_event) {
        next_event = NEXT(event, link);
        task_event_free(&event);
    }
}

void
task_shutdown(task_t task) {
    boolean_t was_idle = FALSE;
    boolean_t discard = FALSE;

    REQUIRE(VALID_TASK(task));

    /*
     * This routine is very similar to task_send_event() above.
     */

    LOCK(&task->lock);
    if (task->enqueue_allowed) {
        if (task->state == task_state_idle) {
            was_idle = TRUE;
            INSIST(EMPTY(task->events));
            task->state = task_state_ready;
        }
        INSIST(task->state == task_state_ready ||
               task->state == task_state_running);
        INSIST(task->shutdown_event != NULL);
        ENQUEUE(task->events, task->shutdown_event, link);
        task->shutdown_event = NULL;
        task->enqueue_allowed = FALSE;
    } else
        discard = TRUE;
    UNLOCK(&task->lock);

    if (discard)
        return;

    if (was_idle) {
        boolean_t need_wakeup = FALSE;
        task_manager_t manager;

        manager = task->manager;
        INSIST(VALID_MANAGER(manager));
        LOCK(&manager->lock);
        if (EMPTY(manager->ready_tasks))
            need_wakeup = TRUE;
        ENQUEUE(manager->ready_tasks, task, ready_link);
        UNLOCK(&manager->lock);

        if (need_wakeup)
            BROADCAST(&manager->work_available);
    }
}

void
task_destroy(task_t *taskp) {

    REQUIRE(taskp != NULL);

    task_shutdown(*taskp);
    task_detach(taskp);
}


/***
 *** Task Manager.
 ***/

static
void *task_manager_run(void *uap) {
    task_manager_t manager = uap;
    task_t task;
    boolean_t no_workers = FALSE;

    XTRACE("start");

    REQUIRE(VALID_MANAGER(manager));

    /*
     * Again we're trying to hold the lock for as short a time as possible
     * and to do as little locking and unlocking as possible.
     *
     * In both while loops, the appropriate lock must be held before the
     * while body starts.  Code which acquired the lock at the top of
     * the loop would be more readable, but would result in a lot of
     * extra locking.  Compare:
     *
     * Straightforward:
     *
     *  LOCK();
     *  ...
     *  UNLOCK();
     *  while (expression) {
     *      LOCK();
     *      ...
     *      UNLOCK();
     *
     *          Unlocked part here...
     *
     *      LOCK();
     *      ...
     *      UNLOCK();
     *  }
     *
     * Note how if the loop continues we unlock and then immediately lock.
     * For N iterations of the loop, this code does 2N+1 locks and 2N+1
     * unlocks.  Also note that the lock is not held when the while
     * condition is tested, which may or may not be important, depending
     * on the expression.
     *
     * As written:
     *
     *  LOCK();
     *  while (expression) {
     *      ...
     *      UNLOCK();
     *
     *          Unlocked part here...
     *
     *      LOCK();
     *      ...
     *  }
     *  UNLOCK();
     *
     * For N iterations of the loop, this code does N+1 locks and N+1
     * unlocks.  The while expression is always protected by the lock.
     */

    LOCK(&manager->lock);
    while (!FINISHED(manager)) {
        /*
         * For reasons similar to those given in the comment in
         * task_send_event() above, it is safe for us to dequeue
         * the task while only holding the manager lock, and then
         * change the task to running state while only holding the
         * task lock.
         */
        while (EMPTY(manager->ready_tasks) && !FINISHED(manager)) {
            XTRACE("wait");
            WAIT(&manager->work_available, &manager->lock);
            XTRACE("awake");
        }
        XTRACE("working");

        task = HEAD(manager->ready_tasks);
        if (task != NULL) {
            unsigned int dispatch_count = 0;
            boolean_t done = FALSE;
            boolean_t requeue = FALSE;
            boolean_t wants_shutdown;
            boolean_t is_shutdown;
            boolean_t free_task = FALSE;
            task_event_t event;
            task_eventlist_t remaining_events;
            boolean_t discard_remaining = FALSE;

            INSIST(VALID_TASK(task));

            /*
             * Note we only unlock the manager lock if we actually
             * have a task to do.  We must reacquire the manager
             * lock before exiting the 'if (task != NULL)' block.
             */
            DEQUEUE(manager->ready_tasks, task, ready_link);
            UNLOCK(&manager->lock);

            LOCK(&task->lock);
            INSIST(task->state == task_state_ready);
            if (EMPTY(task->events)) {
                /*
                 * The task became runnable, but all events
                 * in the run queue were subsequently purged.
                 * Put the task to sleep.
                 */
                task->state = task_state_idle;
                done = TRUE;
                XTRACE("ready but empty");
            } else
                task->state = task_state_running;
            while (!done) {
                INSIST(!EMPTY(task->events));
                event = HEAD(task->events);
                DEQUEUE(task->events, event, link);
                UNLOCK(&task->lock);

                if (event->type == TASK_EVENT_SHUTDOWN)
                    is_shutdown = TRUE;
                else
                    is_shutdown = FALSE;

                /*
                 * Execute the event action.
                 */
                XTRACE("execute action");
                if (event->action != NULL)
                    wants_shutdown =
                        (event->action)(task, event);
                else
                    wants_shutdown = FALSE;
                dispatch_count++;

                task_event_free(&event);

                LOCK(&task->lock);
                if (wants_shutdown || is_shutdown) {
                    /*
                     * The event action has either
                     * requested shutdown, or the event
                     * we just executed was the shutdown
                     * event.
                     *
                     * Since no more events can be
                     * delivered to the task, we purge
                     * any remaining events (but defer
                     * freeing them until we've released
                     * the lock).
                     */
                    XTRACE("wants shutdown");
                    if (!EMPTY(task->events)) {
                        remaining_events =
                            task->events;
                        INIT_LIST(task->events);
                        discard_remaining = TRUE;
                    }
                    if (task->references == 0)
                        free_task = TRUE;
                    task->state = task_state_shutdown;
                    task->enqueue_allowed = FALSE;
                    done = TRUE;
                } else if (EMPTY(task->events)) {
                    /*
                     * Nothing else to do for this task.
                     * Put it to sleep.
                     *
                     * XXX detect tasks with 0 references
                     * and do something about them.
                     */
                    XTRACE("empty");
                    task->state = task_state_idle;
                    done = TRUE;
                } else if (dispatch_count >= task->quantum) {
                    /*
                     * Our quantum has expired, but
                     * there is more work to be done.
                     * We'll requeue it to the ready
                     * queue later.
                     *
                     * We don't check quantum until
                     * dispatching at least one event,
                     * so the minimum quantum is one.
                     */
                    XTRACE("quantum");
                    task->state = task_state_ready;
                    requeue = TRUE;
                    done = TRUE;
                }
            }
            UNLOCK(&task->lock);

            if (discard_remaining) {
                task_event_t next_event;

                for (event = HEAD(remaining_events);
                     event != NULL;
                     event = next_event) {
                    next_event = NEXT(event, link);
                    task_event_free(&event);
                }
            }

            if (free_task)
                task_free(task);

            LOCK(&manager->lock);
            if (requeue) {
                /*
                 * We know we're awake, so we don't have
                 * to wakeup any sleeping threads if the
                 * ready queue is empty before we requeue.
                 *
                 * A possible optimization if the queue is
                 * empty is to 'goto' the 'if (task != NULL)'
                 * block, avoiding the ENQUEUE of the task
                 * and the subsequent immediate DEQUEUE
                 * (since it is the only executable task).
                 * We don't do this because then we'd be
                 * skipping the exit_requested check.  The
                 * cost of ENQUEUE is low anyway, especially
                 * when you consider that we'd have to do
                 * an extra EMPTY check to see if we could
                 * do the optimization.  If the ready queue
                 * were usually nonempty, the 'optimization'
                 * might even hurt rather than help.
                 */
                ENQUEUE(manager->ready_tasks, task,
                    ready_link);
            }
        }
    }
    INSIST(manager->workers > 0);
    manager->workers--;
    if (manager->workers == 0)
        no_workers = TRUE;
    UNLOCK(&manager->lock);

    if (no_workers)
        BROADCAST(&manager->no_workers);

    XTRACE("exit");

    return (NULL);
}

static void
manager_free(task_manager_t manager) {
    (void)os_condition_destroy(&manager->work_available);
    (void)os_condition_destroy(&manager->no_workers);
    (void)os_mutex_destroy(&manager->lock);
    manager->magic = 0;
    mem_put(manager->mctx, manager, sizeof *manager);
}

unsigned int
task_manager_create(mem_context_t mctx, unsigned int workers,
            unsigned int default_quantum, task_manager_t *managerp)
{
    unsigned int i, started = 0;
    task_manager_t manager;
    os_thread_t thread;

    if (workers == 0)
        return (0);
    manager = mem_get(mctx, sizeof *manager);
    if (manager == NULL)
        return (0);
    manager->magic = TASK_MANAGER_MAGIC;
    manager->mctx = mctx;
    if (!os_mutex_init(&manager->lock)) {
        mem_put(mctx, manager, sizeof *manager);
        return (0);
    }
    if (default_quantum == 0)
        default_quantum = DEFAULT_DEFAULT_QUANTUM;
    manager->default_quantum = default_quantum;
    INIT_LIST(manager->tasks);
    INIT_LIST(manager->ready_tasks);
    if (!os_condition_init(&manager->work_available)) {
        (void)os_mutex_destroy(&manager->lock);
        mem_put(mctx, manager, sizeof *manager);
        return (0);
    }
    manager->exiting = FALSE;
    manager->workers = 0;
    if (!os_condition_init(&manager->no_workers)) {
        (void)os_condition_destroy(&manager->work_available);
        (void)os_mutex_destroy(&manager->lock);
        mem_put(mctx, manager, sizeof *manager);
        return (0);
    }

    LOCK(&manager->lock);
    /*
     * Start workers.
     */
    for (i = 0; i < workers; i++) {
        if (os_thread_create(task_manager_run, manager, &thread)) {
            manager->workers++;
            started++;
            (void)os_thread_detach(thread);
        }
    }
    UNLOCK(&manager->lock);

    if (started == 0) {
        manager_free(manager);
        return (0);
    }

    *managerp = manager;

    return (started);
}

void
task_manager_destroy(task_manager_t *managerp) {
    task_manager_t manager;
    task_t task;

    REQUIRE(managerp != NULL);
    manager = *managerp;
    REQUIRE(VALID_MANAGER(manager));

    XTRACE("task_manager_destroy");
    /*
     * Only one non-worker thread may ever call this routine.
     * If a worker thread wants to initiate shutdown of the
     * task manager, it should ask some non-worker thread to call
     * task_manager_destroy(), e.g. by signalling a condition variable
     * that the startup thread is sleeping on.
     */

    /*
     * Unlike elsewhere, we're going to hold this lock a long time.
     * We need to do so, because otherwise the list of tasks could
     * change while we were traversing it.
     *
     * This is also the only function where we will hold both the
     * task manager lock and a task lock at the same time.
     */

    LOCK(&manager->lock);

    /*
     * Make sure we only get called once.
     */
    INSIST(!manager->exiting);
    manager->exiting = TRUE;

    /*
     * Post the shutdown event to every task (if it hasn't already been
     * posted).
     */
    for (task = HEAD(manager->tasks);
         task != NULL;
         task = NEXT(task, link)) {
        LOCK(&task->lock);
        if (task->enqueue_allowed) {
            INSIST(task->shutdown_event != NULL);
            ENQUEUE(task->events, task->shutdown_event, link);
            task->shutdown_event = NULL;
            if (task->state == task_state_idle) {
                task->state = task_state_ready;
                ENQUEUE(manager->ready_tasks, task,
                    ready_link);
            }
            INSIST(task->state == task_state_ready ||
                   task->state == task_state_running);
            task->enqueue_allowed = FALSE;
        }
        UNLOCK(&task->lock);
    }

    /*
     * Wake up any sleeping workers.  This ensures we get work done if
     * there's work left to do, and if there are already no tasks left
     * it will cause the workers to see manager->exiting.
     */
    BROADCAST(&manager->work_available);

    /*
     * Wait for all the worker threads to exit.
     *
     * XXX  This will become a timed wait.  If all the workers haven't
     *      died after we've waited the specified interval, we will
     *  kill the worker threads.  Should we join with the worker
     *      threads after killing them or just leave them detached and
     *      hope they go away?
     */
    while (manager->workers > 0)
        WAIT(&manager->no_workers, &manager->lock);

    UNLOCK(&manager->lock);

    manager_free(manager);

    *managerp = NULL;
}