static
NTSTATUS
EventLoop(
PEPOLL_THREAD pThread
)
{
NTSTATUS status = STATUS_SUCCESS;
RING timed;
RING runnable;
RING waiting;
CLOCK clock = {0};
LONG64 llNow = 0;
LONG64 llNextDeadline = 0;
struct epoll_event events[MAX_EVENTS];
int ready = 0;
BOOLEAN bShutdown = FALSE;
BOOLEAN bSignalled = FALSE;
RingInit(&runnable);
RingInit(&timed);
RingInit(&waiting);
for (;;)
{
/* Get current time for this iteration */
status = ClockGetMonotonicTime(&clock, &llNow);
GOTO_ERROR_ON_STATUS(status);
/* Schedule any timed tasks that have reached their deadline */
ScheduleTimedTasks(
&timed,
llNow,
&runnable);
/* Schedule any waiting tasks that epoll indicated are ready
and check if the thread received a signal */
ScheduleWaitingTasks(
events,
ready,
llNow,
&runnable,
&bSignalled);
if (bSignalled)
{
/* Schedule explicitly-signalled tasks and check if we
have been told to shut down */
ScheduleSignalled(
pThread,
&runnable,
&bShutdown);
}
/* Process runnable tasks */
status = ProcessRunnable(
pThread,
&runnable,
&timed,
&waiting,
llNow);
GOTO_ERROR_ON_STATUS(status);
if (!RingIsEmpty(&runnable))
{
/* If there are still runnable tasks, set the next deadline
to now so we can check for other tasks becoming runnable but
do not block in Poll() */
llNextDeadline = llNow;
}
else if (!RingIsEmpty(&timed))
{
/* There are timed tasks, so set our next deadline to the
deadline of the first task in the queue */
llNextDeadline = LW_STRUCT_FROM_FIELD(timed.pNext, EPOLL_TASK, QueueRing)->llDeadline;
}
else if (!RingIsEmpty(&waiting) || !bShutdown)
{
/* There are waiting tasks or we are not shutting down, so poll indefinitely */
llNextDeadline = -1;
}
else
{
/* We are shutting down and there are no remaining tasks, so leave */
break;
}
/* Wait (or check) for activity */
status = Poll(
&clock,
&llNow,
pThread->EpollFd,
events,
MAX_EVENTS,
llNextDeadline,
&ready);
GOTO_ERROR_ON_STATUS(status);
}
error:
return status;
}
static
NTSTATUS
EventLoop(
PSELECT_THREAD pThread
)
{
NTSTATUS status = STATUS_SUCCESS;
RING tasks;
RING runnable;
PRING pRing = NULL;
PRING pNext = NULL;
PSELECT_TASK pTask = NULL;
CLOCK clock = {0};
LONG64 llNow;
LONG64 llNextDeadline;
fd_set readSet;
fd_set writeSet;
fd_set exceptSet;
int ready = 0;
int nfds = 0;
char c = 0;
int res = 0;
BOOLEAN bShutdown = FALSE;
PLW_TASK_GROUP pGroup = NULL;
BOOLEAN bYielding = FALSE;
RingInit(&tasks);
RingInit(&runnable);
FD_ZERO(&readSet);
FD_ZERO(&writeSet);
FD_ZERO(&exceptSet);
LOCK_THREAD(pThread);
while (!bShutdown || !RingIsEmpty(&tasks))
{
/* Reset variables */
llNextDeadline = 0;
nfds = 0;
bYielding = FALSE;
/* Get current time for this iteration */
GOTO_ERROR_ON_STATUS(status = ClockGetMonotonicTime(&clock, &llNow));
/* Figure out which tasks are runnable */
for (pRing = tasks.pNext; pRing != &tasks; pRing = pNext)
{
pNext = pRing->pNext;
pTask = LW_STRUCT_FROM_FIELD(pRing, SELECT_TASK, EventRing);
/* Update trigger set with results from select() */
UpdateTriggerSet(
pTask,
&readSet,
&writeSet,
&exceptSet,
llNow);
/* Schedule tasks to run if they have been triggered or were yielding */
if ((pTask->TriggerWait & LW_TASK_EVENT_YIELD) ||
((pTask->TriggerWait | LW_TASK_EVENT_EXPLICIT) & pTask->TriggerSet))
{
/* Put task on a separate list to run its trigger function */
RingRemove(&pTask->EventRing);
RingInsertBefore(&runnable, &pTask->EventRing);
/* Update the trigger args with the trigger set */
pTask->TriggerArgs |= pTask->TriggerSet;
/* Turn off bits (except cancel) now that we have copied them */
pTask->TriggerSet &= (LW_TASK_EVENT_CANCEL);
}
else
{
/* Update select parameters to wait for task to trigger */
UpdateTriggerWait(
pTask,
&nfds,
&readSet,
&writeSet,
&exceptSet,
&llNextDeadline);
}
}
UNLOCK_THREAD(pThread);
for (pRing = runnable.pNext; pRing != &runnable; pRing = pNext)
{
pNext = pRing->pNext;
pTask = LW_STRUCT_FROM_FIELD(pRing, SELECT_TASK, EventRing);
GOTO_ERROR_ON_STATUS(status = TaskProcessTrigger(pTask, llNow));
if (pTask->TriggerWait != 0)
{
/* Task is still waiting to be runnable, update select parameters
and put it back in the task list */
UpdateTriggerWait(
pTask,
&nfds,
&readSet,
&writeSet,
&exceptSet,
&llNextDeadline);
if (pTask->TriggerWait & LW_TASK_EVENT_YIELD)
{
/* Task is yielding temporarily. Set the yield flag on
its trigger arguments. Leave it on the runnable list */
pTask->TriggerArgs |= LW_TASK_EVENT_YIELD;
}
else
{
RingRemove(&pTask->EventRing);
RingInsertBefore(&tasks, &pTask->EventRing);
}
}
else
{
/* Task is complete, notify and remove from task group
if it is in one */
RingRemove(&pTask->EventRing);
/* Unregister task from global signal loop */
if (pTask->pUnixSignal)
{
RegisterTaskUnixSignal(pTask, 0, FALSE);
}
pGroup = pTask->pGroup;
if (pGroup)
{
LOCK_GROUP(pGroup);
pTask->pGroup = NULL;
RingRemove(&pTask->GroupRing);
pthread_cond_broadcast(&pGroup->Event);
UNLOCK_GROUP(pGroup);
}
LOCK_THREAD(pThread);
if (--pTask->ulRefCount)
{
pTask->TriggerSet = TASK_COMPLETE_MASK;
pthread_cond_broadcast(&pThread->Event);
UNLOCK_THREAD(pThread);
}
else
{
UNLOCK_THREAD(pThread);
TaskDelete(pTask);
}
}
}
if (!RingIsEmpty(&runnable))
{
/* We have runnable tasks that are yielding. Move them
back to the event list and note the fact. */
bYielding = TRUE;
RingMove(&runnable, &tasks);
}
if (!bShutdown)
{
/* Also wait for a poke on the thread's signal fd */
FD_SET(pThread->SignalFds[0], &readSet);
if (pThread->SignalFds[0] >= nfds)
{
nfds = pThread->SignalFds[0] + 1;
}
}
if (nfds)
{
/* If there are still runnable tasks due to
LW_TASK_EVENT_YIELD, set the next deadline
to now so we wake immediately. This gives other
tasks the chance to become runnable before we
proceed */
if (bYielding)
{
llNextDeadline = llNow;
}
/* Wait for a task to be runnable */
GOTO_ERROR_ON_STATUS(status = Sleep(
&clock,
&llNow,
nfds,
&readSet,
&writeSet,
&exceptSet,
llNextDeadline,
&ready));
}
LOCK_THREAD(pThread);
/* Check for a signal to the thread */
if (FD_ISSET(pThread->SignalFds[0], &readSet))
{
FD_CLR(pThread->SignalFds[0], &readSet);
pThread->bSignalled = FALSE;
res = read(pThread->SignalFds[0], &c, sizeof(c));
assert(res == sizeof(c));
/* Move all tasks in queue into local task list */
RingMove(&pThread->Tasks, &tasks);
if (pThread->bShutdown && !bShutdown)
{
bShutdown = pThread->bShutdown;
/* Cancel all outstanding tasks */
for (pRing = tasks.pNext; pRing != &tasks; pRing = pRing->pNext)
{
pTask = LW_STRUCT_FROM_FIELD(pRing, SELECT_TASK, EventRing);
pTask->TriggerSet |= LW_TASK_EVENT_CANCEL | LW_TASK_EVENT_EXPLICIT;
}
}
}
}
error:
UNLOCK_THREAD(pThread);
return status;
}
static
NTSTATUS
ProcessRunnable(
PEPOLL_THREAD pThread,
PRING pRunnable,
PRING pTimed,
PRING pWaiting,
LONG64 llNow
)
{
NTSTATUS status = STATUS_SUCCESS;
ULONG ulTicks = MAX_TICKS;
PLW_TASK pTask = NULL;
PLW_TASK_GROUP pGroup = NULL;
PRING pRing = NULL;
PRING pNext = NULL;
/* We are guaranteed to run each task at least once. If tasks remain
on the runnable list by yielding, we will continue to run them
all in a round robin until our ticks are depleted. */
while (ulTicks && !RingIsEmpty(pRunnable))
{
for (pRing = pRunnable->pNext; pRing != pRunnable; pRing = pNext)
{
pNext = pRing->pNext;
pTask = LW_STRUCT_FROM_FIELD(pRing, EPOLL_TASK, QueueRing);
RunTask(pTask, llNow);
if (ulTicks)
{
ulTicks--;
}
if (pTask->EventWait != LW_TASK_EVENT_COMPLETE)
{
/* Task is still waiting to be runnable, update events in epoll set */
status = UpdateEventWait(
pTask,
pThread->EpollFd
);
GOTO_ERROR_ON_STATUS(status);
if (pTask->EventWait & LW_TASK_EVENT_YIELD)
{
/* Task is yielding. Set YIELD in its trigger arguments and
and leave it on the runnable list for the next iteration */
pTask->EventArgs |= LW_TASK_EVENT_YIELD;
}
else if (pTask->EventWait & LW_TASK_EVENT_TIME)
{
/* If the task is waiting for a timeout, insert it into the timed queue */
RingRemove(&pTask->QueueRing);
InsertTimedQueue(pTimed, pTask);
}
else
{
/* Otherwise, put it in the generic waiting queue */
RingRemove(&pTask->QueueRing);
RingEnqueue(pWaiting, &pTask->QueueRing);
}
}
else
{
/* Task is complete */
RingRemove(&pTask->QueueRing);
/* Turn off any fd in the epoll set */
if (pTask->Fd >= 0)
{
status = LwRtlSetTaskFd(pTask, pTask->Fd, 0);
GOTO_ERROR_ON_STATUS(status);
}
/* Unsubscribe task from any UNIX signals */
if (pTask->pUnixSignal)
{
RegisterTaskUnixSignal(pTask, 0, FALSE);
}
LOCK_POOL(pThread->pPool);
pThread->ulLoad--;
UNLOCK_POOL(pThread->pPool);
pGroup = pTask->pGroup;
/* If task was in a task group, remove it and notify anyone waiting
on the group */
if (pGroup)
{
LOCK_GROUP(pGroup);
pTask->pGroup = NULL;
RingRemove(&pTask->GroupRing);
pthread_cond_broadcast(&pGroup->Event);
UNLOCK_GROUP(pGroup);
}
LOCK_THREAD(pThread);
if (--pTask->ulRefCount)
{
/* The task still has a reference, so mark it as completed
and notify anyone waiting on it */
pTask->EventSignal = TASK_COMPLETE_MASK;
pthread_cond_broadcast(&pThread->Event);
UNLOCK_THREAD(pThread);
}
else
{
/* We held the last reference to the task, so delete it */
RingRemove(&pTask->SignalRing);
UNLOCK_THREAD(pThread);
TaskDelete(pTask);
}
}
}
}
error:
return status;
}
static
VOID
ProcessRunnable(
PKQUEUE_THREAD pThread,
PKQUEUE_COMMANDS pCommands,
PRING pRunnable,
PRING pTimed,
PRING pWaiting,
LONG64 llNow
)
{
ULONG ulTicks = MAX_TICKS;
PLW_TASK pTask = NULL;
PLW_TASK_GROUP pGroup = NULL;
PRING pRing = NULL;
PRING pNext = NULL;
/* We are guaranteed to run each task at least once. If tasks remain
on the runnable list by yielding, we will continue to run them
all in a round robin until our ticks are depleted. */
while (ulTicks && !RingIsEmpty(pRunnable))
{
for (pRing = pRunnable->pNext; pRing != pRunnable; pRing = pNext)
{
pNext = pRing->pNext;
pTask = LW_STRUCT_FROM_FIELD(pRing, KQUEUE_TASK, QueueRing);
RunTask(pTask, llNow);
if (ulTicks)
{
ulTicks--;
}
if (pTask->EventWait != LW_TASK_EVENT_COMPLETE)
{
if (pTask->EventWait & LW_TASK_EVENT_YIELD)
{
/* Task is yielding. Set the YIELD flag and
leave it on the runnable list for the next iteration. */
pTask->EventArgs |= LW_TASK_EVENT_YIELD;
}
else
{
/* Task is still waiting on events, update kqueue */
UpdateEventWait(pCommands, pTask);
if (pTask->EventWait & LW_TASK_EVENT_TIME)
{
/* If the task is waiting for a timeout,
insert it into the timed queue */
RingRemove(&pTask->QueueRing);
InsertTimedQueue(pTimed, pTask);
}
else
{
/* Otherwise, put it in the generic waiting queue */
RingRemove(&pTask->QueueRing);
RingEnqueue(pWaiting, &pTask->QueueRing);
}
}
}
else
{
/* Task is complete */
RingRemove(&pTask->QueueRing);
/* Remove any associated events from the kqueue */
if (pTask->Fd >= 0)
{
(void) LwRtlSetTaskFd(pTask, pTask->Fd, 0);
}
/* Unsubscribe task from any UNIX signals */
if (pTask->pUnixSignal)
{
RegisterTaskUnixSignal(pTask, 0, FALSE);
}
LOCK_POOL(pThread->pPool);
pThread->ulLoad--;
UNLOCK_POOL(pThread->pPool);
pGroup = pTask->pGroup;
/* If task was in a task group, remove it and notify anyone waiting
on the group */
if (pGroup)
{
LOCK_GROUP(pGroup);
pTask->pGroup = NULL;
RingRemove(&pTask->GroupRing);
pthread_cond_broadcast(&pGroup->Event);
UNLOCK_GROUP(pGroup);
}
LOCK_THREAD(pThread);
if (--pTask->ulRefCount)
{
/* The task still has a reference, so mark it as completed
and notify anyone waiting on it */
pTask->EventSignal = TASK_COMPLETE_MASK;
pthread_cond_broadcast(&pThread->Event);
UNLOCK_THREAD(pThread);
}
else
{
/* We held the last reference to the task, so delete it */
RingRemove(&pTask->SignalRing);
UNLOCK_THREAD(pThread);
TaskDelete(pTask);
}
}
}
}
/* Update kevent commands for yielding tasks */
for (pRing = pRunnable->pNext; pRing != pRunnable; pRing = pRing->pNext)
{
pTask = LW_STRUCT_FROM_FIELD(pRing, KQUEUE_TASK, QueueRing);
if (pTask->EventArgs & LW_TASK_EVENT_YIELD)
{
UpdateEventWait(pCommands, pTask);
}
}
}