/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.cassandra.concurrent;
import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.ConcurrentLinkedQueue;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicLong;
import org.apache.cassandra.metrics.SEPMetrics;
import org.apache.cassandra.utils.concurrent.SimpleCondition;
import org.apache.cassandra.utils.concurrent.WaitQueue;
import static org.apache.cassandra.concurrent.SEPWorker.Work;
public class SEPExecutor extends AbstractLocalAwareExecutorService
{
private final SharedExecutorPool pool;
public final int maxWorkers;
public final String name;
private final int maxTasksQueued;
private final SEPMetrics metrics;
// stores both a set of work permits and task permits:
// bottom 32 bits are number of queued tasks, in the range [0..maxTasksQueued] (initially 0)
// top 32 bits are number of work permits available in the range [0..maxWorkers] (initially maxWorkers)
private final AtomicLong permits = new AtomicLong();
// producers wait on this when there is no room on the queue
private final WaitQueue hasRoom = new WaitQueue();
private final AtomicLong completedTasks = new AtomicLong();
volatile boolean shuttingDown = false;
final SimpleCondition shutdown = new SimpleCondition();
// TODO: see if other queue implementations might improve throughput
protected final ConcurrentLinkedQueue<FutureTask<?>> tasks = new ConcurrentLinkedQueue<>();
SEPExecutor(SharedExecutorPool pool, int maxWorkers, int maxTasksQueued, String jmxPath, String name)
{
this.pool = pool;
this.name = name;
this.maxWorkers = maxWorkers;
this.maxTasksQueued = maxTasksQueued;
this.permits.set(combine(0, maxWorkers));
this.metrics = new SEPMetrics(this, jmxPath, name);
}
protected void onCompletion()
{
completedTasks.incrementAndGet();
}
// schedules another worker for this pool if there is work outstanding and there are no spinning threads that
// will self-assign to it in the immediate future
boolean maybeSchedule()
{
if (pool.spinningCount.get() > 0 || !takeWorkPermit(true))
return false;
pool.schedule(new Work(this));
return true;
}
protected void addTask(FutureTask<?> task)
{
// we add to the queue first, so that when a worker takes a task permit it can be certain there is a task available
// this permits us to schedule threads non-spuriously; it also means work is serviced fairly
tasks.add(task);
int taskPermits;
while (true)
{
long current = permits.get();
taskPermits = taskPermits(current);
// because there is no difference in practical terms between the work permit being added or not (the work is already in existence)
// we always add our permit, but block after the fact if we breached the queue limit
if (permits.compareAndSet(current, updateTaskPermits(current, taskPermits + 1)))
break;
}
if (taskPermits == 0)
{
// we only need to schedule a thread if there are no tasks already waiting to be processed, as
// the original enqueue will have started a thread to service its work which will have itself
// spawned helper workers that would have either exhausted the available tasks or are still being spawned.
// to avoid incurring any unnecessary signalling penalties we also do not take any work to hand to the new
// worker, we simply start a worker in a spinning state
pool.maybeStartSpinningWorker();
}
else if (taskPermits >= maxTasksQueued)
{
// register to receive a signal once a task is processed bringing the queue below its threshold
WaitQueue.Signal s = hasRoom.register();
// we will only be signalled once the queue drops below full, so this creates equivalent external behaviour
// however the advantage is that we never wake-up spuriously;
// we choose to always sleep, even if in the intervening time the queue has dropped below limit,
// so long as we _will_ eventually receive a signal
if (taskPermits(permits.get()) > maxTasksQueued)
{
// if we're blocking, we might as well directly schedule a worker if we aren't already at max
if (takeWorkPermit(true))
pool.schedule(new Work(this));
metrics.totalBlocked.inc();
metrics.currentBlocked.inc();
s.awaitUninterruptibly();
metrics.currentBlocked.dec();
}
else // don't propagate our signal when we cancel, just cancel
s.cancel();
}
}
// takes permission to perform a task, if any are available; once taken it is guaranteed
// that a proceeding call to tasks.poll() will return some work
boolean takeTaskPermit()
{
while (true)
{
long current = permits.get();
int taskPermits = taskPermits(current);
if (taskPermits == 0)
return false;
if (permits.compareAndSet(current, updateTaskPermits(current, taskPermits - 1)))
{
if (taskPermits == maxTasksQueued && hasRoom.hasWaiters())
hasRoom.signalAll();
return true;
}
}
}
// takes a worker permit and (optionally) a task permit simultaneously; if one of the two is unavailable, returns false
boolean takeWorkPermit(boolean takeTaskPermit)
{
int taskDelta = takeTaskPermit ? 1 : 0;
while (true)
{
long current = permits.get();
int workPermits = workPermits(current);
int taskPermits = taskPermits(current);
if (workPermits == 0 || taskPermits == 0)
return false;
if (permits.compareAndSet(current, combine(taskPermits - taskDelta, workPermits - 1)))
{
if (takeTaskPermit && taskPermits == maxTasksQueued && hasRoom.hasWaiters())
hasRoom.signalAll();
return true;
}
}
}
// gives up a work permit
void returnWorkPermit()
{
while (true)
{
long current = permits.get();
int workPermits = workPermits(current);
if (permits.compareAndSet(current, updateWorkPermits(current, workPermits + 1)))
return;
}
}
public void maybeExecuteImmediately(Runnable command)
{
FutureTask<?> ft = newTaskFor(command, null);
if (!takeWorkPermit(false))
{
addTask(ft);
}
else
{
try
{
ft.run();
}
finally
{
returnWorkPermit();
// we have to maintain our invariant of always scheduling after any work is performed
// in this case in particular we are not processing the rest of the queue anyway, and so
// the work permit may go wasted if we don't immediately attempt to spawn another worker
maybeSchedule();
}
}
}
public synchronized void shutdown()
{
shuttingDown = true;
pool.executors.remove(this);
if (getActiveCount() == 0)
shutdown.signalAll();
// release metrics
metrics.release();
}
public synchronized List<Runnable> shutdownNow()
{
shutdown();
List<Runnable> aborted = new ArrayList<>();
while (takeTaskPermit())
aborted.add(tasks.poll());
return aborted;
}
public boolean isShutdown()
{
return shuttingDown;
}
public boolean isTerminated()
{
return shuttingDown && shutdown.isSignaled();
}
public boolean awaitTermination(long timeout, TimeUnit unit) throws InterruptedException
{
shutdown.await(timeout, unit);
return isTerminated();
}
public long getPendingTasks()
{
return taskPermits(permits.get());
}
public long getCompletedTasks()
{
return completedTasks.get();
}
public int getActiveCount()
{
return maxWorkers - workPermits(permits.get());
}
private static int taskPermits(long both)
{
return (int) both;
}
private static int workPermits(long both)
{
return (int) (both >>> 32);
}
private static long updateTaskPermits(long prev, int taskPermits)
{
return (prev & (-1L << 32)) | taskPermits;
}
private static long updateWorkPermits(long prev, int workPermits)
{
return (((long) workPermits) << 32) | (prev & (-1L >>> 32));
}
private static long combine(int taskPermits, int workPermits)
{
return (((long) workPermits) << 32) | taskPermits;
}
}